Skip to content
Agent Fundamentals

Execute

Execute is the agent’s part. Expand left behind a task list — typically “n tools to implement for this platform” — and a TOOL.md contract in each tool folder. The agent now writes the platform file beside each contract:

.agent/tools/check-plist-keys/
├── TOOL.md        ← the contract (came from the source)
├── fixtures/      ← test files the contract's examples refer to
└── macos.py       ← written here, by the agent, for this machine

The contract fixes everything that matters — input schema, output schema, declared effects, worked examples — so “implement this” is a tightly bounded task, not an invitation to be creative:

  • The wire format is non-negotiable: one JSON object on stdin, one JSON object on stdout, exit 0 when ok is true.
  • The language is free. On this machine it was python3 (macos.py) for plist parsing and a shell script (macos.sh) where codesign does the real work. Another platform gets its own file — windows.ps1 next to macos.py — and both are committed, so the next machine of the same platform implements nothing.
  • Error cases are part of the contract. If the spec’s examples include a missing-file case returning ok: false with an error string, the implementation must do exactly that.

The iteration rule: implement, then run speccify tool check — and keep going until every example passes. The agent does not declare a tool done; the check does.

Once the tools stand, executing the skill is nothing special: the agent follows the expanded SKILL.md step by step, calling the tools it now trusts, exactly as described under Skills. Speccify adds only a trace — which skill ran, which iteration, when — so that later there is an answer to “what did it actually do?”.

Whether the result is any good is the question Evaluate answers.