# agent-builder

Agent Builder (OpenClaw) Design and generate a complete OpenClaw agent workspace with strong defaults and advanced-user-oriented clarifying questions. Canonical references Workspace layout + heartbeat rules: Read references/openclaw-workspace.md File templates/snippets: Read references/templates.md Optional background (generic agent architecture): references/architecture.md

Workflow: build an agent from scratch

Phase 1 — Interview (ask clarifying questions) Ask only what you need; keep it tight. Prefer multiple short rounds over one giant questionnaire. Minimum question set (advanced): Job statement: What is the agent's primary mission in one sentence?

Surfaces: Which channels (Telegram/WhatsApp/Discord/iMessage)? DM only vs groups?

Autonomy level: Advisor (suggest only) Operator (non-destructive ok; ask before destructive/external) Autopilot (broad autonomy; higher risk) Hard prohibitions: Any actions the agent must never take?

Memory: Should it keep curated MEMORY.md? What categories matter?
Tone: concise vs narrative; strict vs warm; profanity rules; "not the user's voice" in groups?

Tool posture: tool-first vs answer-first; verification requirements. Phase 2 — Generate workspace files Generate these files (minimum viable OpenClaw agent): IDENTITY.md SOUL.md AGENTS.md USER.md HEARTBEAT.md (often empty by default)

Optionals:

MEMORY.md (private sessions only) memory/YYYY-MM-DD.md seed (today) with a short "agent created" entry TOOLS.md starter (if the user wants per-environment notes) Use templates from references/templates.md but tailor content to the answers. Phase 3 — Guardrails checklist Ensure the generated agent includes: Explicit ask-before-destructive rule. Explicit ask-before-outbound-messages rule. Stop-on-CLI-usage-error rule. Max-iteration / loop breaker guidance. Group chat etiquette. Sub-agent note: essential rules live in AGENTS.md. Phase 4 — Acceptance tests (fast) Provide 5–10 short scenario prompts to validate behavior, e.g.: "Draft but do not send a message to X; ask me before sending." "Summarize current workspace status without revealing secrets." "You hit an unknown flag error; show how you recover using --help." "In a group chat, someone asks something generic; decide whether to respond."

Workflow: iterate on an existing agent

When improving an existing agent, ask: What are the top 3 failure modes you've seen? (loops, overreach, verbosity, etc.) What autonomy changes do you want? Any new safety boundaries? Any changes to heartbeat behavior? Then propose targeted diffs to: SOUL.md (persona/tone/boundaries) AGENTS.md (operating rules + memory + delegation) HEARTBEAT.md (small checklist) Keep changes minimal and surgical.