Skip to content

The Durability Curve · free by email

Get the Weekly every Saturday, with the calls we grade, plus each new piece.

107Pieces, dated
1245Claims verified at the primary source
7Instruments, no login

Your address goes to Substack, where you confirm it. Nothing is stored on this site.

The latest issue · Anthropic Says It Has Met Its Own AI Research Threshold. NVIDIA's New Chief Challenged Huang's Bets. Read it first

Start here: Start with the problem you have.

The Durability Curve is Harry Floyd’s publication on how AI changes what work is worth.

  • Run one thing on your own work
  • Read the piece to start with
  • Check the evidence, claim by claim
The idea in one picture: what lasts loses its value slowly, and the surface loses it fast.

01 · What are you working on?

5 problems

Deciding whether to trust an agent

You are about to put an autonomous system somewhere it can do damage. Five pieces, in the order that builds the check.

  1. How often a job finishes, by its length, at two per‑step success rates.

    RunFree, no login

    The Marathon Calculator

    An agent that gets each step right 95% of the time finishes a 40-step job 13% of the time. At 98% it finishes 45%.

    Run it: The Marathon Calculator

  2. The path’s pieces in minutes. This one is first.

    Read11 min

    How Reliable Is Your AI Agent?

    A month running an autonomous agent. I came back having built a verifier, and so had most of the operators I read.

    Read it: How Reliable Is Your AI Agent?

  3. One mark a claim. verified at the source.

    CheckClaim ledger

    The Guardrail Your Agent Can Reach

    A companion piece on the same problem: every claim, with its source, and how far it was checked.

    See the evidence: The Guardrail Your Agent Can Reach

Then, in order4 more · 43 min

  1. You’re Not Comparing Models. You’re Comparing Contracts. 9 min
  2. Your AI Looks Best Where You Can Check It Least 10 min
  3. Most Verification Is Just Bigger Classification 16 min
  4. Never Let Claude Code Tell You It’s Done 8 min
The whole path on one page: Deciding whether to trust an agent

The Seven-Layer Agent AuditFor your team: a two-page PDF, no sign-up Download The Seven-Layer Agent Audit (PDF)

Why your benchmark is lying to you

Every number on your dashboard was produced by something. This is what each one is actually measuring.

  1. Stays green while its meaning sinks

    RunFree, no login

    The Metric Validity Audit

    Name the number you steer by. You get the specific way it can mislead you, and the check that catches it.

    Run it: The Metric Validity Audit

  2. The path’s pieces in minutes. This one is first.

    Read11 min

    Ten Lines of Code Scored 100%. One Agent Broke Eight Benchmarks.

    Not one task was actually solved, and the same blind spot is sitting in your own dashboard.

    Read it: Ten Lines of Code Scored 100%. One Agent Broke Eight Benchmarks.

  3. One mark a claim. verified at the source, not yet.

    CheckClaim ledger

    Your AI Grader Is Only as Good as Its Answer Key

    A companion piece on the same problem: every claim, with its source, and how far it was checked.

    See the evidence: Your AI Grader Is Only as Good as Its Answer Key

Then, in order4 more · 36 min

  1. A Green Score Is Not Evidence 8 min
  2. The Stable Liar 11 min
  3. Where Your Metrics Fold 7 min
  4. Your Benchmark Measures a Sprint. Your Agent Runs a Marathon. 10 min
The whole path on one page: Why your benchmark is lying to you

Building an agent that gets better

The model is the part you cannot change. Five pieces on the parts you can.

  1. A team of agents folded into one loop

    RunFree, no login

    The Multi-Agent Decision

    List your tasks and answer four questions on each. You get a ranked call on which ones one agent in a loop handles better than a team of agents.

    Run it: The Multi-Agent Decision

  2. The path’s pieces in minutes. This one is first.

    Read11 min

    Same Model, Different Product: The Case for Harness Engineering

    Hold the model fixed and the surrounding harness can still move task success, cost and reliability. Here are the layers that do it.

    Read it: Same Model, Different Product: The Case for Harness Engineering

  3. One mark a claim. verified at the source, not yet.

    CheckClaim ledger

    The Limit Said 10. The Loop Made 500 Calls.

    A companion piece on the same problem: every claim, with its source, and how far it was checked.

    See the evidence: The Limit Said 10. The Loop Made 500 Calls.

Then, in order4 more · 36 min

  1. Your AI Agent Stack Is Solving The Wrong Problem 11 min
  2. Remembers Everything, Learns Nothing 11 min
  3. Self-Improvement Is Release Engineering 7 min
  4. Skills Are Package Management for Your AI 7 min
The whole path on one page: Building an agent that gets better

The Skill Bill of MaterialsFor your team: a two-page PDF, no sign-up Download The Skill Bill of Materials (PDF)

Where the money actually lands

Everyone can see the buildout. Far fewer can say which layer of it keeps the return.

  1. Your pace vs the layer getting cheaper

    RunFree, no login

    The Two-Rate Diagnostic

    Two questions give you a dated window: roughly when your AI advantage stops paying, and which layer to build on next.

    Run it: The Two-Rate Diagnostic

  2. The path’s pieces in minutes. This one is first.

    Read10 min

    How Long Until Your AI Edge Stops Paying?

    You adopted AI everywhere and it still didn't pay. The scarce layer keeps the money, until its clock runs out.

    Read it: How Long Until Your AI Edge Stops Paying?

  3. One mark a claim. verified at the source.

    CheckClaim ledger

    Amazon Shut Out Meta’s Agent. Shopify Gave It Shop Pay.

    A companion piece on the same problem: every claim, with its source, and how far it was checked.

    See the evidence: Amazon Shut Out Meta's Agent. Shopify Gave It Shop Pay.

Then, in order4 more · 48 min

  1. The Forest Floor Is The Product 20 min
  2. Three Hidden Bottlenecks the AI Buildout Has Already Moved Past GPUs 11 min
  3. The Other Half of Compute 9 min
  4. Right About AI, Wiped Out Anyway 8 min
The whole path on one page: Where the money actually lands

What the machine is doing to you

The capability arrived. Five pieces on what it quietly took on the way in.

  1. Help on, then help off: one trial

    Run9 min

    Use AI to Learn Without Getting Worse at It

    Answer-giving AI left learners worse off once it was gone. Three setups built to keep the thinking with you, and a way to check what stuck.

    Open the runbook: Use AI to Learn Without Getting Worse at It

  2. The path’s pieces in minutes. This one is first.

    Read16 min

    The Difficulty You’re Escaping Was Making You

    AI can lift the effort out of almost anything you find hard. Some of that effort was the thing turning you into someone.

    Read it: The Difficulty You're Escaping Was Making You

  3. One mark a claim. verified at the source, not yet.

    CheckClaim ledger

    Confidently Wrong

    Every claim in this piece, with its source, and how far it was checked.

    See the evidence: Confidently Wrong

Then, in order4 more · 42 min

  1. Confidently Wrong 10 min
  2. The Safe Parts of Your Job Are the First to Go 10 min · part paid
  3. Taste Is What You Delete 9 min
  4. What Proves You Can Think? 13 min
The whole path on one page: What the machine is doing to you

Know what you want? All 7 tools Every piece, by path The essay map

Get the next piece by email

02 · The lens

Two minutes

What survives when the surface changes?

Every piece here asks it of whatever looks impressive. Here it is, with a test you can run on your own work.

The surfaceLast 30 days
  1. Google Gemini 4 Carbon
  2. Anthropic Claude Fable 5.2
  3. Mistral Mistral Large 4 (open weights)
  4. OpenAI GPT-6.1 Terra
  5. Moonshot Kimi K3.1 Pro
  6. DeepSeek DeepSeek-V4.1-Flash-1017
  7. Meta Muse Spark Open
  8. Google Gemini 4 Argon (GA)

8 frontier releases in 30 days 177 since Nov 2022

Each row links to the lab’s own announcement. From the log behind the AI release chart, data through .

The surfacehere called canopy

  • this month’s model
  • the demo
  • the benchmark score
  • the prompt that works this week

Repriced every few weeks

What lastshere called substrate

  • the tests you trust
  • your data
  • the workflow it plugs into
  • the record of what broke
  • the rule you decide by
  • the trust you’ve earned

Still has a job after the repricing

Where the value goes nexthere called the bottleneck

To the neighbouring layer that is now hardest to get: up to whoever checks and chooses, or down to power, chips and permission. Something built to last still loses if the value moves away from it.

The test, in three moves

  1. Commit

    Write down what you think will last, before anything happens.

  2. Name the change

    Be exact: a cheaper open model, a channel that stops favouring you, a platform that ships your feature.

  3. Count

    What still has a job is how durable it was.

Commit before you name the change. Do it the other way round and whatever survived will look like the lasting part.

A worked case, at my own expense

I wrote a script to run the seven questions automatically, then threw it away, because a script reads your file names rather than your setup and returns a confident verdict on any stack it does not recognise. The automation was canopy. The questions were substrate. I had built the wrong one first.

Harry Floyd, on building The Seven-Layer Agent Audit

Three questions to carry

  1. What is the part that lasts here?
  2. What happens to it when the ground moves?
  3. Is the value moving toward it, or away?

The long versionThe original Start Here essay, 6 min

03 · The record

1564 claims · 36 ledgers

How to check what you read here

Who writes it

I’m Harry Floyd. I run a research system that turns AI, product, and market noise into instruments for seeing what survives when the surface changes.

About the publication

The newsletter · Free

Subscribe to get each new piece in your inbox, with something to use in every one.

Several essays a week, plus Notes through the week.

17 new pieces here in the 30 days to 15 Nov 2026. 90 of the 107 are free in full.

Already subscribed? Every piece is in the archive, or follow by RSS.

Signup completes on Substack. Unsubscribe from any email.