Build agents that hold up
Put effort where it moves the agent, give it the context and memory that change what it does, and ship changes you can roll back. Each skill is taught by the essays here and tested on real cases. You answer first, say how sure you are, then see what happened and the source.
01 · The map
7 skills · lines join a skill to the ones it needs
The skills below are in map order. Each one says which skills it needs first.
- Proved on your own work
- Held: right again after a gap
- Passed
- Started
- No piece teaches it yet
02 · The skills
Read the pieces, then practise the cases
-
1
Spend effort where it moves the agent most
Most agent work goes into the parts that matter least. Find the layer that is actually starving it.
- The Leverage Hierarchy of Agent Engineering
- Your AI Agent Stack Is Solving The Wrong Problem
- The Seven-Layer Agent Audit
-
2
Write the file your agent reads first
The context file decides more than the prompt. Treat it as a spec, score it, and fix its thinnest part.
-
3
Make memory change what the agent does
Remembering is cheap. An agent learns only when a lesson is tested and changes its next run.
Needs first: 2 · Write the file your agent reads first
-
4
Ship a change to an agent like a release
An agent that edits its own memory or skills has shipped a change. See what changed, test it, and keep a way back.
Needs first: 3 · Make memory change what the agent does
-
5
Choose and package skills you can trust
A skill file is a dependency. Know where it came from, what it can touch, and whether it ever runs.
Needs first: 2 · Write the file your agent reads first
- Skills Are Package Management for Your AI
- The Skill That Never Fired
-
6
Keep agents from losing each other's decisions
Several agents share state like an org chart shares decisions. Find where a verdict can be overwritten with no error.
Needs first: 1 · Spend effort where it moves the agent most
- Your Multi-Agent System Is an Org Chart
- The Graph Looks Right. The Merge Is Where It Breaks.
-
7
Spot the old structure behind a new AI problem
A benchmark, a model jury and an agent swarm are old structures with known failures. Recognise them before you rebuild the fix.
Needs first: 1 · Spend effort where it moves the agent most