agentlisted

# Synthesize a bounded agent Stated need: `$ARGUMENTS` Order matters: **need → spec → validate → emit.** Writing an agent into `.claude/agents/` first and running it second would turn a definition an agent produced into one every future session inherits. This skill declares no pre-approved tools. Promotion writes a file that grants standing authority, so it goes through the normal permission flow, deliberately. ## Preconditions If `scripts/ai-harness/harness_agentgen.py` does not exist, stop and say so: the harness is Lite, or none is installed. ## Resolve the Python interpreter ```bash python3 --version ``` If that fails, use `python --version`. Substitute the name that printed a version for `<python>` below. ## 1. Ask whether it should exist at all An existing agent that nearly fits beats a new one. Read `.claude/agents/` first. A need that is really one task is a session (`/development-harness:session`), not a new standing agent. ## 2. Write the need ```json { "name": "<kebab-case-name>", "need": "<one sentence: what nobody currently knows or does>", "capability": "reader", "duties": ["<verifiable duty>", "<verifiable duty>"] } ``` `reader` is the default and the right answer far more often than it feels. Choose `verifier` only when the agent must run something, and `implementer` only when it must write — which additionally requires `writable_paths` and `approved_by_operator: true`, and the operator's approval must be real, not assumed from the reques