Skip to content

What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes

Lessons from OpenAI, Anthropic, Google and independent engineers, worked through one failure, with a way to test every change.

Claims
85
Verified at the primary source
84
Removed before publication
0
Last checked

85 claims: 84 verified, 1 executed.

Download this ledger: CSV · JSON · Markdown · CC BY 4.0 · Query it from Claude or ChatGPT

What the essay claims

Much of an agent's reliability comes from what surrounds the model (instructions, context, tools, checks, sign-off points), and the published guides agree on a short set of steps. But every addition to a harness costs something and encodes an assumption, and the only controlled tests of instruction files found no reliable success gain. So a harness is built one tested fix at a time, against the simpler setup it replaces. The reader starts believing either that reliability is a model-choice problem, or that more rules always help.

The claim ladder

What the ladder does not reach: none is a controlled comparison of harness against model; no step is shown to transfer to every tool or codebase (OpenAI says its results depend on its repo's structure); the token multipliers (4× / 15×) are Anthropic's own data on its research system. The guide reports what the sources did and recommends; it measures nothing itself.

#ClaimWhose behaviour it measuresStatus / scope
1Harness engineering is designing everything around the model so the agent gets work right first time and fixes its own mistakesDefinition, synthesised from Böckeler, Hashimoto, ThoughtworksDefinitional; no measurement
2Shorter, hand-written instruction files work better than long or auto-generated onesOpenAI's team, HumanLayer, Software Mansion (practitioner reports)Reported practice, not a controlled test
3Context quality falls as context grows; clearing, compaction and notes each trade something awayAnthropic (engineering post + cookbook)Vendor guidance
4Curated memory raised an agent's pass rate 22.3% → 68.4% with 31% fewer tokensDatabricks' own system on Databricks' internal benchmarkVendor benchmark; one system, one benchmark
5Silent passing checks and agent-readable error messages keep agents on taskHumanLayer (4,000-line incident), OpenAI (custom lint messages), ThoughtworksPractitioner reports
6A feature list + progress file + commit per session lets agents finish long tasksAnthropic's long-running agents experiment (one test app)One vendor experiment
7Behavioural evals on one failure mode at a time, read as a trendGoogle's Gemini CLI teamVendor method
8Context files did not significantly change task success (generated: marginal negative; developer-written: marginal gain) and raised cost by over 20%ETH Zurich (Gloaguen et al.), Python repos, several agents and modelsControlled study; Python only; not significant either way
9An AGENTS.md cut median runtime 28.6% and output tokens 16.6% with comparable completionLulla et al., 10 repos, 124 PRsControlled study, small; efficiency only
10Every harness component encodes an assumption about what the model can't do, which can go stale as models improveAnthropic's harness-design team (own experience)Vendor guidance
11Tool search defers tool definitions until needed (Anthropic: ~85% fewer tokens in their example; Claude Code defers MCP tools)Anthropic docs and engineering postVendor documentation
12Hooks run with the user's full permissions; deny rules and sandboxing are separate layersClaude Code docsVendor documentation

What would make it wrong

The guide is wrong if a reader who follows the checklist (one-screen instructions file, silent checks with fix-carrying errors, tests the agent cannot edit, a done-list for long tasks) sees no change in how often their agent declares unfinished work done.

On the track record, not yet testable1 untestedThe re-check

The evidence, row by row

Each row is a claim the essay makes, then the evidence recorded for it and the source it was checked against. For verified rows that is usually the exact words, where they sit and the day they were read. The label says how far it was checked.

What the labels mean
Verified
Checked against the primary source itself by a checker who did not draft the piece, usually with the exact words, where they sit and the date read.
Executed
A run or observation done for the piece: code, a query, a count, or a check made live in an app. The claim is what it returned. Run records are not published.
Checked
Checked against its source by a separate checker in the older ledger format, which recorded no quote, locator or date.
Reported
Not verified word for word against a primary source: carried from a secondary source, from one that could not be opened in full, or supported only in part. The essay words it accordingly.
Struck
Drafted, checked and removed before publication. Kept here, crossed out, with the reason.
Excluded
Considered and deliberately left out of the piece. Kept here with the reason.
Sources (20)
  1. Thoughtworks 8 rows
  2. Böckeler 6 rows
  3. Hashimoto 1 row
  4. Anthropic 17 rows
  5. SmartScope 2 rows
  6. agents.md 1 row
  7. Claude Code docs 5 rows
  8. OpenAI 7 rows
  9. HumanLayer 8 rows
  10. Software Mansion 5 rows
  11. Claude Cookbook 2 rows
  12. Sourcegraph 5 rows
  13. Databricks 2 rows
  14. MCP specification 2025-06-18 2 rows
  15. The Durability Curve 1 row
  16. Google 6 rows
  17. Gloaguen et al. 4 rows
  18. Lulla et al. 1 row
  19. Vitest docs 1 row
  20. Our run 1 row
  1. 1 Verified

    Thoughtworks describes an agent renaming a field that two downstream services consume; calls it locally correct but systemically destructive

    "locally correct, but systemically destructive"

    Thoughtworks (Nandyala, Pani), 29 Sep 2026 · opening section · thoughtworks.com

    Link
    Cite

    Thoughtworks (Nandyala, Pani), 29 Sep 2026, opening section, https://www.thoughtworks.com/insights/blog/architecture/engineering-the-harness-a-practical-pattern-for-reliable-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 1, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-1.

    Report an error
  2. 2 Verified

    Thoughtworks says the gap is not mainly about model quality

    "is not primarily a model-quality issue"

    Thoughtworks · opening section

    Link
    Cite

    Thoughtworks, opening section. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 2, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-2.

    Report an error
  3. 3 Verified

    Agent = Model + Harness

    "Agent = Model + Harness"

    Böckeler, martinfowler.com, 2 Apr 2026 (also Thoughtworks) · early section · martinfowler.com

    Link
    Cite

    Böckeler, martinfowler.com, 2 Apr 2026 (also Thoughtworks), early section, https://martinfowler.com/articles/harness-engineering.html. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 3, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-3.

    Report an error
  4. 4 Verified

    Part of the harness is built into the tool (system prompt, code retrieval); users build an outer harness for their use case

    "outer harness"

    Böckeler · "Metaphors only go so far" area

    Link
    Cite

    Böckeler, "Metaphors only go so far" area. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 4, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-4.

    Report an error
  5. 5 Verified

    A good outer harness raises first-time success and gives a self-correcting feedback loop

    "self-corrects as many issues as possible before they even reach human eyes"

    Böckeler · after Figure 1

    Link
    Cite

    Böckeler, after Figure 1. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 5, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-5.

    Report an error
  6. 6 Verified

    Hashimoto's definition of harness engineering; he says in this post he had started calling the practice harness engineering

    "is the idea that anytime you find an agent makes a mistake, you take the time to engineer a solution such that the agent never makes that mistake again"

    Hashimoto, My AI Adoption Journey, 5 Feb 2026 · Step 5 · mitchellh.com

    Link
    Cite

    Hashimoto, My AI Adoption Journey, 5 Feb 2026, Step 5, https://mitchellh.com/writing/my-ai-adoption-journey. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 6, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-6.

    Report an error
  7. 7 Verified

    Anthropic calls context engineering the next step after prompt engineering

    "the natural progression of prompt engineering"

    Anthropic, 29 Sep 2025 · intro · anthropic.com

    Link
    Cite

    Anthropic, 29 Sep 2025, intro, https://www.anthropic.com/research/effective-context-engineering-for-ai-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 7, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-7.

    Report an error
  8. 8 Verified

    Context engineering targets one output; harness engineering targets quality over time

    (paraphrase of comparison table: "A single output is inaccurate" vs "Quality degrades over time")

    SmartScope, Feb 2026, restructured Mar 2026 · comparison table · smartscope.blog

    Link
    Cite

    SmartScope, Feb 2026, restructured Mar 2026, comparison table, https://smartscope.blog/en/blog/harness-engineering-overview/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 8, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-8.

    Report an error
  9. 9 Verified

    AGENTS.md is a shared format; nearest file wins in nested repos

    "so the closest one takes precedence"

    agents.md · FAQ / monorepo · agents.md

    Link
    Cite

    agents.md, FAQ / monorepo, https://agents.md/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 9, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-9.

    Report an error
  10. 10 Verified

    Claude Code reads AGENTS.md only when no CLAUDE.md exists (default)

    "By default, Claude reads AGENTS.md only when you have no CLAUDE.md in your working directory or above it"

    Claude Code docs · When Claude Code reads AGENTS.md · code.claude.com

    Link
    Cite

    Claude Code docs, When Claude Code reads AGENTS.md, https://code.claude.com/docs/en/memory. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 10, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-10.

    Report an error
  11. 11 Verified

    OpenAI tried one big AGENTS.md and it failed

    "one big AGENTS.md"

    OpenAI (Lopopolo), 11 Feb 2026 · We made repository knowledge the system of record · openai.com

    Link
    Cite

    OpenAI (Lopopolo), 11 Feb 2026, We made repository knowledge the system of record, https://openai.com/index/harness-engineering/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 11, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-11.

    Report an error
  12. 12 Verified

    A giant file crowds out the task

    "crowds out the task"

    OpenAI · same

    Link
    Cite

    OpenAI, same. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 12, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-12.

    Report an error
  13. 13 Verified

    Too much guidance becomes non-guidance

    "When everything is 'important,' nothing is."

    OpenAI · same

    Link
    Cite

    OpenAI, same. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 13, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-13.

    Report an error
  14. 14 Verified

    OpenAI's AGENTS.md is about 100 lines and acts as a table of contents pointing to docs/

    "roughly 100 lines"; "the table of contents"

    OpenAI · same

    Link
    Cite

    OpenAI, same. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 14, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-14.

    Report an error
  15. 15 Verified

    Map, not manual

    "give Codex a map, not a 1,000-page instruction manual"

    OpenAI · same

    Link
    Cite

    OpenAI, same. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 15, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-15.

    Report an error
  16. 16 Verified

    What the agent can't access in context doesn't exist for it

    "anything it can't access in-context while running effectively doesn't exist"

    OpenAI · Agent legibility is the goal

    Link
    Cite

    OpenAI, Agent legibility is the goal. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 16, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-16.

    Report an error
  17. 17 Verified

    HumanLayer's CLAUDE.md is under 60 lines; craft by hand, don't auto-generate

    "under 60 lines"

    HumanLayer (Kyle), 12 Mar 2026 · CLAUDE.md section · hlyr.dev

    Link
    Cite

    HumanLayer (Kyle), 12 Mar 2026, CLAUDE.md section, https://www.hlyr.dev/blog/skill-issue-harness-engineering-for-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 17, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-17.

    Report an error
  18. 18 Verified

    /init output brings little value

    "tends to produce documents that bring little meaningful value"

    Software Mansion, updated 22 Jun 2026 · AGENTS.md · agentic-engineering.swmansion.com

    Link
    Cite

    Software Mansion, updated 22 Jun 2026, AGENTS.md, https://agentic-engineering.swmansion.com/becoming-productive/harness-engineering/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 18, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-18.

    Report an error
  19. 19 Verified

    Put-in / leave-out table adapted from Anthropic's Claude Code docs, as quoted by Software Mansion

    (adapted table)

    Software Mansion · AGENTS.md: good and bad practices

    Link
    Cite

    Software Mansion, AGENTS.md: good and bad practices. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 19, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-19.

    Report an error
  20. 20 Verified

    Recall falls as context grows

    "the model's ability to accurately recall information from that context decreases"

    Anthropic context engineering · context rot · anthropic.com

    Link
    Cite

    Anthropic context engineering, context rot, https://www.anthropic.com/research/effective-context-engineering-for-ai-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 20, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-20.

    Report an error
  21. 21 Verified

    Aim for the smallest high-signal set

    "the smallest possible set of high-signal tokens"

    Anthropic context engineering · same

    Link
    Cite

    Anthropic context engineering, same. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 21, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-21.

    Report an error
  22. 22 Verified

    Three techniques: tool-result clearing, compaction, notes/memory; clearing swaps old results for a placeholder and the agent can re-fetch

    (paraphrase)

    Claude Cookbook (Isabella He), 20 Mar 2026 · summary table / clearing · platform.claude.com

    Link
    Cite

    Claude Cookbook (Isabella He), 20 Mar 2026, summary table / clearing, https://platform.claude.com/cookbook/tool-use-context-engineering-context-engineering-tools. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 22, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-22.

    Report an error
  23. 23 Verified

    Clearing is a light, safe form of compaction

    "one of the safest lightest touch forms of compaction"

    Anthropic context engineering · compaction · anthropic.com

    Link
    Cite

    Anthropic context engineering, compaction, https://www.anthropic.com/research/effective-context-engineering-for-ai-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 23, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-23.

    Report an error
  24. 24 Verified

    Sourcegraph's compaction prompt preserves task, files inspected, decisions, failed approaches

    "approaches already tried that failed, and why"; "and the fresh window will cheerfully retry them"

    Sourcegraph (Tanner), 15 Sep 2026 · compaction prompt · sourcegraph.com

    Link
    Cite

    Sourcegraph (Tanner), 15 Sep 2026, compaction prompt, https://sourcegraph.com/blog/context-engineering. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 24, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-24.

    Report an error
  25. 25 Verified

    Databricks' own memory system (MemCurrent) on its internal benchmark of enterprise tasks: 22.3% (no memory) → 68.4% pass rate, 31% fewer tokens, 26% faster than no memory; Mem0 reached 43.9% on the same benchmark

    "more history is not necessarily better memory"

    Databricks, 20 Oct 2026 · intro + Figure 6 · databricks.com

    Link
    Cite

    Databricks, 20 Oct 2026, intro + Figure 6, https://www.databricks.com/blog/memcurrent-designing-persistent-memory-production-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 25, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-25.

    Report an error
  26. 26 Verified

    A stale memory is worse than none

    "is worse than no memory: the agent holds a confident, wrong prior it will defend"

    Sourcegraph · memory invalidation (JWT example) · sourcegraph.com

    Link
    Cite

    Sourcegraph, memory invalidation (JWT example), https://sourcegraph.com/blog/context-engineering. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 26, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-26.

    Report an error
  27. 27 Verified

    Mechanism rule of thumb (AGENTS.md, skills, MCP, subagents, hooks)

    (paraphrased table)

    Software Mansion · When to use what · agentic-engineering.swmansion.com

    Link
    Cite

    Software Mansion, When to use what, https://agentic-engineering.swmansion.com/becoming-productive/harness-engineering/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 27, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-27.

    Report an error
  28. 28 Verified

    Role-based subagents don't work; use them for context control

    "It doesn't work."

    HumanLayer · Sub-Agents Are for Context Control · hlyr.dev

    Link
    Cite

    HumanLayer, Sub-Agents Are for Context Control, https://www.hlyr.dev/blog/skill-issue-harness-engineering-for-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 28, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-28.

    Report an error
  29. 29 Verified

    Subagents may use tens of thousands of tokens and return 1,000-2,000

    "tens of thousands of tokens"; "often 1,000-2,000 tokens"

    Anthropic context engineering · sub-agent architectures · anthropic.com

    Link
    Cite

    Anthropic context engineering, sub-agent architectures, https://www.anthropic.com/research/effective-context-engineering-for-ai-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 29, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-29.

    Report an error
  30. 30 Verified

    Too many tools is bad

    (heading "Too Many Tools Is Bad")

    HumanLayer · MCP Servers section · hlyr.dev

    Link
    Cite

    HumanLayer, MCP Servers section, https://www.hlyr.dev/blog/skill-issue-harness-engineering-for-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 30, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-30.

    Report an error
  31. 31 Verified

    Ambiguous tool choice defeats agents too

    "If a human engineer can't definitively say which tool should be used in a given situation, an AI agent can't be expected to do better."

    Anthropic context engineering · tools · anthropic.com

    Link
    Cite

    Anthropic context engineering, tools, https://www.anthropic.com/research/effective-context-engineering-for-ai-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 31, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-31.

    Report an error
  32. 32 Verified

    Don't connect untrusted MCP servers

    "never connect to one you don't trust. This can be a dangerous vector for prompt injection!"

    HumanLayer · MCP warning · hlyr.dev

    Link
    Cite

    HumanLayer, MCP warning, https://www.hlyr.dev/blog/skill-issue-harness-engineering-for-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 32, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-32.

    Report an error
  33. 33 Verified

    MCP spec: clients must treat tool annotations as untrusted unless from trusted servers

    "clients MUST consider tool annotations to be untrusted unless they come from trusted servers"

    MCP specification 2025-06-18 · Tool definition · modelcontextprotocol.io

    Link
    Cite

    MCP specification 2025-06-18, Tool definition, https://modelcontextprotocol.io/specification/2025-06-18/server/tools. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 33, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-33.

    Report an error
  34. 34 Verified

    Published attacks: a skill step fetched and ran ransomware; hidden Unicode instructions in a skill

    "visual source review alone may miss malicious behavior"

    Software Mansion · Skills: Security considerations · agentic-engineering.swmansion.com

    Link
    Cite

    Software Mansion, Skills: Security considerations, https://agentic-engineering.swmansion.com/becoming-productive/harness-engineering/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 34, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-34.

    Report an error
  35. 35 Verified

    HumanLayer auto-denies migration commands and tells the agent to ask the user

    (paraphrase: "we automatically deny any Bash() tool calls that try to run migrations, with an instruction to ask the user to run them instead")

    HumanLayer · Hooks

    Link
    Cite

    HumanLayer, Hooks. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 35, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-35.

    Report an error
  36. 36 Verified

    Guides steer before, sensors check after

    "Guides (feedforward controls)"; "Sensors (feedback controls)"

    Böckeler · Feedforward and Feedback · martinfowler.com

    Link
    Cite

    Böckeler, Feedforward and Feedback, https://martinfowler.com/articles/harness-engineering.html. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 36, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-36.

    Report an error
  37. 37 Verified

    Guides alone / sensors alone failure modes

    "that encodes rules but never finds out whether they worked"; "that keeps repeating the same mistakes"

    Böckeler · same

    Link
    Cite

    Böckeler, same. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 37, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-37.

    Report an error
  38. 38 Verified

    The rule is called silent success, verbose failure

    "silent success, verbose failure"

    Thoughtworks · Sensors · thoughtworks.com

    Link
    Cite

    Thoughtworks, Sensors, https://www.thoughtworks.com/insights/blog/architecture/engineering-the-harness-a-practical-pattern-for-reliable-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 38, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-38.

    Report an error
  39. 39 Verified

    Full test suite output flooded context; the agent lost track

    "4,000 lines of passing tests would flood the context window"

    HumanLayer · Verification · hlyr.dev

    Link
    Cite

    HumanLayer, Verification, https://www.hlyr.dev/blog/skill-issue-harness-engineering-for-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 39, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-39.

    Report an error
  40. 40 Verified

    Custom lint messages carry fixes

    "inject remediation instructions into agent context"

    OpenAI · Enforcing architecture and taste · openai.com

    Link
    Cite

    OpenAI, Enforcing architecture and taste, https://openai.com/index/harness-engineering/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 40, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-40.

    Report an error
  41. 41 Verified

    Böckeler calls this positive prompt injection

    "a positive kind of prompt injection"

    Böckeler · sensors · martinfowler.com

    Link
    Cite

    Böckeler, sensors, https://martinfowler.com/articles/harness-engineering.html. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 41, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-41.

    Report an error
  42. 42 Verified

    Repeated prose rules should become sensors

    "should be promoted into a mechanical sensor"

    Thoughtworks · sensors · thoughtworks.com

    Link
    Cite

    Thoughtworks, sensors, https://www.thoughtworks.com/insights/blog/architecture/engineering-the-harness-a-practical-pattern-for-reliable-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 42, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-42.

    Report an error
  43. 43 Verified

    Instructions for judgement, hooks for the rest

    "Written instructions should cover the cases that require judgment. Hooks should cover the cases that do not."

    Software Mansion · Hooks · agentic-engineering.swmansion.com

    Link
    Cite

    Software Mansion, Hooks, https://agentic-engineering.swmansion.com/becoming-productive/harness-engineering/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 43, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-43.

    Report an error
  44. 44 Verified

    In Anthropic's long-running setup, later sessions could only flip a feature-list item's passes flag, under an instruction not to remove or edit tests (the 'tests' are feature-list entries)

    "unacceptable to remove or edit tests"

    Anthropic long-running agents, 26 Nov 2025 · Feature list · anthropic.com

    Link
    Cite

    Anthropic long-running agents, 26 Nov 2025, Feature list, https://www.anthropic.com/research/effective-harnesses-for-long-running-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 44, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-44.

    Report an error
  45. 45 Verified

    Each new session starts with no memory; engineers-in-shifts analogy

    "begins with no memory of what came before"

    Anthropic long-running · The long-running agent problem

    Link
    Cite

    Anthropic long-running, The long-running agent problem. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 45, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-45.

    Report an error
  46. 46 Verified

    Two failure modes: doing too much at once; later agents declaring the job done

    "to do too much at once"; "declare the job done"

    Anthropic long-running · same

    Link
    Cite

    Anthropic long-running, same. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 46, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-46.

    Report an error
  47. 47 Verified

    Initializer writes 200+ features, all marked failing; JSON chosen over Markdown

    "less likely to inappropriately change or overwrite JSON files compared to Markdown"

    Anthropic long-running · Feature list

    Link
    Cite

    Anthropic long-running, Feature list. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 47, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-47.

    Report an error
  48. 48 Verified

    Progress file + git commits allow reverting bad changes; start-up routine reads git log and progress, picks top feature, runs a basic test

    (paraphrase)

    Anthropic long-running · Incremental progress; Getting up to speed

    Link
    Cite

    Anthropic long-running, Incremental progress; Getting up to speed. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 48, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-48.

    Report an error
  49. 49 Verified

    Testing as a human user caught bugs not visible in code

    "as a human user would"; "that weren't obvious from the code alone"

    Anthropic long-running · Testing

    Link
    Cite

    Anthropic long-running, Testing. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 49, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-49.

    Report an error
  50. 50 Verified

    Human gates for irreversible or high-blast-radius choices (multi-repo scope, schema migrations, public API changes); otherwise sensible defaults

    "irreversible or high-blast-radius choices"; "sensible defaults"

    Thoughtworks · Guides: human gates · thoughtworks.com

    Link
    Cite

    Thoughtworks, Guides: human gates, https://www.thoughtworks.com/insights/blog/architecture/engineering-the-harness-a-practical-pattern-for-reliable-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 50, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-50.

    Report an error
  51. 51 Verified

    Least-privilege tools: verification agents get read-only tools

    (paraphrase: "provisioned with read-only tools, literally lacking file-writing capabilities")

    Thoughtworks · Least-privilege tools

    Link
    Cite

    Thoughtworks, Least-privilege tools. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 51, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-51.

    Report an error
  52. 52 Verified

    Before You Trust Your AI Agent kit is free, seven checks, one page each

    "Seven checks, one page each"

    The Durability Curve, 9 Oct 2026 · subtitle · harryfloyd.substack.com

    Link
    Cite

    The Durability Curve, 9 Oct 2026, subtitle, https://harryfloyd.substack.com/p/before-you-trust-your-ai-agent. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 52, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-52.

    Report an error
  53. 53 Verified

    Scores move a few points with no idea why

    "by a few percentage points, and have no idea why it changed"

    Google Developers Blog (Mullen, Gunderman), 9 Sep 2026 · intro · developers.googleblog.com

    Link
    Cite

    Google Developers Blog (Mullen, Gunderman), 9 Sep 2026, intro, https://developers.googleblog.com/the-anatomy-of-harness-engineering-how-to-evaluate-iterate-and-guard-ai-coding-agents/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 53, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-53.

    Report an error
  54. 54 Verified

    Behavioural evals check discrete actions (clarifying question, validator before done)

    (paraphrase)

    Google · behavioural eval examples

    Link
    Cite

    Google, behavioural eval examples. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 54, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-54.

    Report an error
  55. 55 Verified

    Loop: pick one failure mode; strict vs LLM-judge checks; batch runs because single runs are noisy

    (paraphrase)

    Google · three-step loop

    Link
    Cite

    Google, three-step loop. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 55, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-55.

    Report an error
  56. 56 Verified

    Suite purpose

    "did not make the agent holistically worse"

    Google · When to evaluate

    Link
    Cite

    Google, When to evaluate. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 56, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-56.

    Report an error
  57. 57 Verified

    Log context on failed runs and check for the answer

    "then check whether the answer was in there"

    Sourcegraph · observability · sourcegraph.com

    Link
    Cite

    Sourcegraph, observability, https://sourcegraph.com/blog/context-engineering. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 57, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-57.

    Report an error
  58. 58 Verified

    Cost scales with what you send

    "Cost and latency scale with what you send, not with what you could send"

    Sourcegraph · cost

    Link
    Cite

    Sourcegraph, cost. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 58, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-58.

    Report an error
  59. 59 Verified

    Agents ~4x chat tokens, multi-agent ~15x

    "agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats"

    Anthropic, 13 Jun 2025 · benefits and downsides · anthropic.com

    Link
    Cite

    Anthropic, 13 Jun 2025, benefits and downsides, https://www.anthropic.com/research/multi-agent-research-system. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 59, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-59.

    Report an error
  60. 60 Verified

    Earn every rule

    "Earn every rule"

    Thoughtworks · close · thoughtworks.com

    Link
    Cite

    Thoughtworks, close, https://www.thoughtworks.com/insights/blog/architecture/engineering-the-harness-a-practical-pattern-for-reliable-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 60, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-60.

    Report an error
  61. 61 Verified

    HumanLayer discards most hooks

    "I have thrown away many more hooks than we actually use today"

    HumanLayer · Hooks · hlyr.dev

    Link
    Cite

    HumanLayer, Hooks, https://www.hlyr.dev/blog/skill-issue-harness-engineering-for-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 61, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-61.

    Report an error
  62. 62 Verified

    ETH Zurich study: context files (LLM-generated and developer-committed) did not generally improve task success and raised inference cost by over 20%

    "does not generally improve task success rates, while increasing inference cost by over 20% on average"

    Gloaguen et al., arXiv 2602.11988, v3 29 Sep 2026 · Abstract · arxiv.org

    Link
    Cite

    Gloaguen et al., arXiv 2602.11988, v3 29 Sep 2026, Abstract, https://arxiv.org/abs/2602.11988. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 62, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-62.

    Report an error
  63. 63 Verified

    Repository overviews not helpful; files useful for non-standard practices

    "repository overviews, although popular and recommended by model providers, are not helpful"; "useful for specifying non-standard coding practices"

    Gloaguen et al. · Abstract · arxiv.org

    Link
    Cite

    Gloaguen et al., Abstract, https://arxiv.org/abs/2602.11988. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 63, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-63.

    Report an error
  64. 64 Verified

    Every harness component encodes an assumption about what the model can't do, and these go stale as models improve

    "encodes an assumption about what the model can't do on its own"; "can quickly go stale as models improve"

    Anthropic, Harness design for long-running application development, 24 Mar 2026 · simplifying the harness · anthropic.com

    Link
    Cite

    Anthropic, Harness design for long-running application development, 24 Mar 2026, simplifying the harness, https://www.anthropic.com/research/harness-design-long-running-apps. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 64, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-64.

    Report an error
  65. 65 Verified

    HumanLayer frames harness engineering as a subset of context engineering, via harness configuration points

    "the subset of context engineering which primarily involves leveraging harness configuration points"

    HumanLayer · What is harness engineering · hlyr.dev

    Link
    Cite

    HumanLayer, What is harness engineering, https://www.hlyr.dev/blog/skill-issue-harness-engineering-for-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 65, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-65.

    Report an error
  66. 66 Verified

    Tool-result clearing, compaction and memory are offered as API features (cookbook audience = API builders); Claude Code uses compaction and memory

    "Our API offers first-party implementations of all three"

    Claude Cookbook · intro · platform.claude.com

    Link
    Cite

    Claude Cookbook, intro, https://platform.claude.com/cookbook/tool-use-context-engineering-context-engineering-tools. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 66, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-66.

    Report an error
  67. 67 Verified

    The comparison table's question row and 'typical failure' row come from SmartScope's context-vs-harness table

    (paraphrase: "What to show" vs "What to prevent, measure, control, and fix"; "A single output is inaccurate" vs "Quality degrades over time")

    SmartScope · comparison table · smartscope.blog

    Link
    Cite

    SmartScope, comparison table, https://smartscope.blog/en/blog/harness-engineering-overview/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 67, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-67.

    Report an error
  68. 68 Verified

    Lulla et al.: with AGENTS.md, median runtime −28.64%, output tokens −16.58%, comparable task completion (10 repos, 124 PRs)

    "lower median runtime"; "reduced output token consumption"; "comparable task completion behavior"

    Lulla et al., arXiv 2601.20404, 28 Jan 2026 · Abstract · arxiv.org

    Link
    Cite

    Lulla et al., arXiv 2601.20404, 28 Jan 2026, Abstract, https://arxiv.org/abs/2601.20404. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 68, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-68.

    Report an error
  69. 69 Verified

    Thoughtworks: what the model lacked in the rename case was system-level visibility, guardrails and automated feedback

    "system-level visibility, architectural guardrails and automated feedback mechanisms"

    Thoughtworks · opening · thoughtworks.com

    Link
    Cite

    Thoughtworks, opening, https://www.thoughtworks.com/insights/blog/architecture/engineering-the-harness-a-practical-pattern-for-reliable-coding-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 69, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-69.

    Report an error
  70. 70 Verified

    Google runs evals in batches because single runs are noisy

    "can be noisy due to nondeterminism of AI models"

    Google Developers Blog · batch evaluations · developers.googleblog.com

    Link
    Cite

    Google Developers Blog, batch evaluations, https://developers.googleblog.com/the-anatomy-of-harness-engineering-how-to-evaluate-iterate-and-guard-ai-coding-agents/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 70, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-70.

    Report an error
  71. 71 Verified

    Anthropic's working rule, from Building Effective Agents

    "find the simplest solution possible, and only increase complexity when needed"

    Anthropic harness design · simplifying the harness · anthropic.com

    Link
    Cite

    Anthropic harness design, simplifying the harness, https://www.anthropic.com/research/harness-design-long-running-apps. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 71, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-71.

    Report an error
  72. 72 Verified

    Anthropic's Tool Search Tool defers tool definitions until needed; ~85% fewer tokens in its example

    "85% reduction in token usage"; "defer_loading"

    Anthropic, Introducing advanced tool use, 24 Nov 2025 · Tool Search Tool · anthropic.com

    Link
    Cite

    Anthropic, Introducing advanced tool use, 24 Nov 2025, Tool Search Tool, https://www.anthropic.com/research/advanced-tool-use. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 72, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-72.

    Report an error
  73. 73 Verified

    Claude Code defers MCP tool definitions until needed

    "deferring tool definitions until Claude needs them"

    Claude Code docs, MCP · Scale with MCP tool search · code.claude.com

    Link
    Cite

    Claude Code docs, MCP, Scale with MCP tool search, https://code.claude.com/docs/en/mcp. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 73, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-73.

    Report an error
  74. 74 Verified

    Claude Code hooks run shell commands with the user's full permissions

    "execute shell commands with your full user permissions"

    Claude Code docs, hooks · Security considerations · code.claude.com

    Link
    Cite

    Claude Code docs, hooks, Security considerations, https://code.claude.com/docs/en/hooks. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 74, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-74.

    Report an error
  75. 75 Verified

    Claude Code deny rules block a tool and are evaluated first

    "Deny rules prevent Claude Code from using the specified tool"; "Rules are evaluated in order: deny, then ask, then allow"

    Claude Code docs, permissions · permission rules · code.claude.com

    Link
    Cite

    Claude Code docs, permissions, permission rules, https://code.claude.com/docs/en/permissions. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 75, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-75.

    Report an error
  76. 76 Verified

    Claude Code's sandbox limits which directories a command can write to

    "restricts file system access to specific directories"

    Claude Code docs, sandboxing · How sandboxing works · code.claude.com

    Link
    Cite

    Claude Code docs, sandboxing, How sandboxing works, https://code.claude.com/docs/en/sandboxing. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 76, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-76.

    Report an error
  77. 77 Verified

    Databricks: episodic memory (raw session traces, searchable) alone 22.3%→35.8%; derived memory 66.1%; both 68.4%; internal benchmark of enterprise tasks

    "Episodic memory alone helps moderately: 35.8%, up from 22.3% with no memory"; "lifting pass rate to 66.1%"; "best result, at 68.4%"

    Databricks · What Makes MemCurrent Work · databricks.com

    Link
    Cite

    Databricks, What Makes MemCurrent Work, https://www.databricks.com/blog/memcurrent-designing-persistent-memory-production-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 77, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-77.

    Report an error
  78. 78 Verified

    A later long-running agent would see progress and declare the job done

    "look around, see that progress had been made, and declare the job done"

    Anthropic long-running · The long-running agent problem · anthropic.com

    Link
    Cite

    Anthropic long-running, The long-running agent problem, https://www.anthropic.com/research/effective-harnesses-for-long-running-agents. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 78, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-78.

    Report an error
  79. 79 Verified

    ETH measured inference cost; Lulla measured runtime and output tokens (different efficiency measures)

    "increasing inference cost by over 20% on average"; "lower median runtime"; "reduced output token consumption"

    Gloaguen et al.; Lulla et al. · Abstracts · arxiv.org

    Link
    Cite

    Gloaguen et al.; Lulla et al., Abstracts, https://arxiv.org/abs/2602.11988. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 79, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-79.

    Report an error
  80. 80 Verified

    Sourcegraph warns a badly-retrieving sub-agent still returns a confident summary

    "a sub-agent that retrieved badly still returns a confident summary"

    Sourcegraph · Sub-agent architectures · sourcegraph.com

    Link
    Cite

    Sourcegraph, Sub-agent architectures, https://sourcegraph.com/blog/context-engineering. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 80, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-80.

    Report an error
  81. 81 Verified

    An MCP server's list of tools can change after a client has read it (list-changed notification)

    "When the list of available tools changes, servers that declared the listChanged capability SHOULD send a notification"

    MCP specification 2025-06-18 · List Changed Notification · modelcontextprotocol.io

    Link
    Cite

    MCP specification 2025-06-18, List Changed Notification, https://modelcontextprotocol.io/specification/2025-06-18/server/tools. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 81, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-81.

    Report an error
  82. 82 Verified

    Google distinguishes end-to-end benchmarks (outcomes) from behavioural evals (discrete actions)

    "End-to-end benchmarks are the de facto for evaluating model performance"; "a behavioral eval measures discrete, observable actions"

    Google Developers Blog · intro; behavioural evals · developers.googleblog.com

    Link
    Cite

    Google Developers Blog, intro; behavioural evals, https://developers.googleblog.com/the-anatomy-of-harness-engineering-how-to-evaluate-iterate-and-guard-ai-coding-agents/. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 82, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-82.

    Report an error
  83. 83 Verified

    Vitest accepts a failure message as expect's second argument (used in the code example)

    "Vitest supports a message as the second argument"

    Vitest docs, expect · expect · vitest.dev

    Link
    Cite

    Vitest docs, expect, expect, https://vitest.dev/api/expect. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 83, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-83.

    Report an error
  84. 84 Verified

    ETH: developer-provided files did slightly better, 2.4 percentage points on average, not significant (p=21%), and significantly outperformed LLM-generated files (p=3.8%)

    "Developer-provided context files improve agent performance by 2.4% on average (p=21%), significantly outperforming LLM-generated ones (p=3.8%)"

    Gloaguen et al. (full paper, §4.2) · 4.2 Main results · arxiv.org

    Link
    Cite

    Gloaguen et al. (full paper, §4.2), 4.2 Main results, https://arxiv.org/html/2602.11988v3. Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 84, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-84.

    Report an error
  85. 85 Executed

    The guide's contract check fails on the renamed schema with the quoted message and passes on the original

    Result: (run output)

    Our run, 22 Oct 2026, Vitest 5.0.5 + Zod + TypeScript 7.0.2

    A run or observation done for this piece. The run record is not published.

    Link
    Cite

    A run or observation by The Durability Curve (Our run, 22 Oct 2026, Vitest 5.0.5 + Zod + TypeScript 7.0.2). Checked in: Claim Ledger, "What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes", The Durability Curve, row 85, https://durabilitycurve.com/claims/what-is-harness-engineering/#c-85.

    Report an error

Cite and reuse

A single claim
Use the row's Cite button. It names the original source first, then this row, which carries its own link.
This ledger
Floyd, Harry (2026). Claim Ledger: What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes. The Durability Curve. https://durabilitycurve.com/claims/what-is-harness-engineering/
The essay
Floyd, Harry (2026). What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes. The Durability Curve. https://durabilitycurve.com/blog/what-is-harness-engineering/

The ledger data is licensed CC BY 4.0: reuse it, quote it or train on it, with credit to The Durability Curve and a link. Read the licence. Machine copies: JSON, CSV, Markdown.

The newsletter · Free

Subscribe to get the next structural lens in your inbox.

Several essays a week, plus Notes through the week.

Signup completes on Substack. Unsubscribe from any email.