agent-builder

agent-builder — build any goal-declared agent

This skill is the meta-system that generalizes the proven diligence agent ("Atlas") into a reusable pipeline: declare a goal, and out comes a portable, self-hardening, publishable agent. Atlas shipped to prod using the patterns this skill reuses; agent-builder turns those patterns from "build one diligence agent" into "build any agent". You build the builder, not the agent's output.

Explicit goal

Take a user's declared goal ("an agent that does X") and produce a portable git repo on the OpenClaw/Claude-Code harness that achieves it, ships its own adversarial safeguards by default, survives a fresh no-memory publish gauntlet twice with zero CRITICALs before it is allowed to ship, and serializes to a .af agent-file for handover. Refuse to declare an agent publishable until the gate says so.

Lifecycle

[1] DECLARE   — resolve the goal to ONE testable sentence. Ask 2-3 clarifying Qs
                only if it is too vague to test. Write it to agent.config.json.
[2] CHOOSE    — pick runtime/model/tools + write policies/hooks/safeguards.
                The decision table below drives every choice from the goal.
[3] SCAFFOLD  — scripts/scaffold_agent.py emits the portable repo (CLAUDE.md,
                .claude/skills/ incl. the self-adversarial critic BY DEFAULT,
                .claude/hooks/, .mcp.json, policies/, tests/, manifest).
[4] BUILD     — flesh out the domain skills/prompts so it achieves the goal;
                keep the self-critic wired.
[5] RUN       — exercise the agent; produce a sample deliverable.
[6] GATE      — scripts/publish_gate.py: a fresh no-memory adversarial gauntlet.
                quality sublayer ALWAYS; ai-detection sublayer when opted in.
                Counter resets on any failed round.
[7] ITERATE   — apply the cited fixes; re-run. One pass never unlocks.
[8] PUBLISH   — ONLY after two consecutive zero-CRITICAL rounds (publishable:true).
[9] EXPORT    — scripts/af_export.py emits the .af handover artifact.