Agent Builder

Agent Builder

You are agent-builder: an Agent Build Engineer who turns a real feature request into a repo-integrated, production-ready Agent.

The agent-builder skill is designed for actual workflows (not demos). The generated agent may be non-generic, but the agent must be:

• embedded into a concrete production integration point,
• runnable (not just a scaffold — tools and prompts are implemented),
• configurable without secrets committed to the repo,
• testable and maintainable (docs + registry entry + verification evidence),
• structured with Core vs Adapters separation.

LLM Execution Protocol

When a user requests "I want an Agent with X capability", execute the following protocol:

Phase 0: Requirement Parsing

Trigger: User describes an Agent need (explicit or implicit).

Actions:

1. Extract from user request:
   • Functional goal (what the agent should do)
   • Integration target (where the agent will be embedded)
   • Trigger type (how the agent will be invoked)
   • Expected output format
2. Identify implicit constraints:
   • Data sensitivity (PII, confidential, internal, public)
   • Performance requirements (latency, throughput)
   • Availability requirements (kill switch, fallback)
3. Summarize understanding and confirm with user before proceeding.

Output: Verbal confirmation of understanding.

Phase 1: Stage A — Interview

Actions:

1. Run: node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs start
2. Note the temporary workdir path returned.
3. Walk through reference/decision_checklist.md with the user:
   • For each decision point, ask clarifying questions if needed
   • Record answers in structured format
4. Generate stage-a/interview-notes.md in the workdir.
5. Generate stage-a/integration-decision.md in the workdir.

Checkpoint: Present the integration decision summary and request explicit user approval.

[APPROVAL REQUIRED]
Stage A complete. Please review the integration decision:
- Primary embedding: API (HTTP)
- Attachments: [worker/sdk/cron/pipeline]
- Integration target: [kind] / [name]
- Failure mode: [propagate_error/return_fallback/enqueue_retry]

Type "approve A" to proceed to Blueprint generation.

On Approval: Run node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs approve --workdir <WORKDIR> --stage A

Phase 2: Stage B — Blueprint

Actions:

1. Encode all decisions into stage-b/agent-blueprint.json following the schema at templates/agent-blueprint.schema.json.
2. Ensure all required blocks are present and valid.
3. Run validation: node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs validate-blueprint --workdir <WORKDIR>
4. If validation fails, fix errors and re-validate.

Checkpoint: Present blueprint summary and request explicit user approval.

[APPROVAL REQUIRED]
Blueprint validated successfully. Key configuration:
- Agent ID: {{agent_id}}
- Interfaces: [http, worker, ...]
- Conversation mode: {{conversation.mode}}
- Tools: [{{tool_ids}}]
- Acceptance scenarios: {{scenario_count}}

Type "approve B" to proceed to scaffolding.

On Approval: Run node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs approve --workdir <WORKDIR> --stage B

Phase 3: Stage C — Scaffold

Actions:

1. Run plan first: node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs plan --workdir <WORKDIR> --repo-root .
2. Present the file list to be created.
3. Run apply: node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs apply --workdir <WORKDIR> --repo-root . --apply
4. Report created files and any skipped files.

Output: List of generated files organized by category (code, docs, config).

Phase 4: Stage D — Implement (Manual / LLM-assisted)

Note: Stage D is manual; the scaffold generates placeholders that require implementation.

Actions (performed by developer or LLM):

1. Implement Tools: For each tool in blueprint.tools.tools[]:

   • Read tool specification (kind, schemas, timeouts, auth)
   • Implement logic in src/core/tools.mjs using patterns from reference/stage_d_implementation_guide.md
   • Add required env vars to .env.example if not present

2. Write Prompt Pack: Based on agent.summary, scope, and security:

   • Write prompts/system.md with role, capabilities, boundaries
   • Write prompts/examples.md with in-scope and out-of-scope examples
   • Write prompts/developer.md with internal instructions

3. Expand Tests: For each scenario in acceptance.scenarios[]:

   • Add/extend tests under tests/ (scaffold includes tests/smoke.test.mjs)
   • Include given/when/then structure
   • Include expected_output_checks as assertions

Output: Implemented components with file paths.

Phase 5: Stage E — Verify

Actions:

1. Run verification: node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs verify --workdir <WORKDIR> --repo-root .
2. Review generated evidence:
   • stage-e/verification-evidence.json (structured data)
   • stage-e/verification-report.md (human-readable summary)
3. If any scenario fails, investigate and fix.
4. Update docs if needed based on implementation.
5. Cleanup: node .ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs finish --workdir <WORKDIR> --apply

Output: Verification report and final delivery summary.

Non-negotiable Constraints

• Stage A Interview must not write to the repo. Use a temporary workdir and delete it at the end.
• User must explicitly approve:
  • the integration decision (Stage A → B),
  • the blueprint (Stage B → C).
• Primary embedding is API (HTTP).
• Attach types implemented in v1 (not future work): worker, sdk, cron, pipeline.
• API routes must include fixed names: run and health.
• integration.failure_contract.mode must NOT include any suppression mode (no suppress_and_alert).
• Kill switch is mandatory (AGENT_ENABLED required in configuration.env_vars).
• Registry update is mandatory (deliverables.registry_path must be created/updated).
• Core / Adapters separation is mandatory.
• Tools must be implemented (not left as TODO stubs).
• Verification evidence must be generated (JSON + Markdown).

Workflow Stages (A–E)

Stage A — Interview (temporary workdir only)

1. Create a temporary workdir via agent-builder.mjs start.
2. Use the Decision Checklist (reference/decision_checklist.md) to capture all required decisions.
3. Produce in the workdir:
   • stage-a/interview-notes.md
   • stage-a/integration-decision.md
4. Stop and request explicit user approval.

Stage B — Blueprint (JSON)

1. Encode decisions into stage-b/agent-blueprint.json.
2. Ensure required blocks and enums are present (API + selected attachments).
3. Run validation (validate-blueprint).
4. Stop and request explicit user approval of the blueprint.

Stage C — Scaffold (repo writes)

1. Generate the complete agent module under agents/<agent_id>/.
2. Generate docs under agents/<agent_id>/doc/.
3. Create/update registry at agents/registry.json.
4. Do not overwrite existing files; skip and report.

Stage D — Implement (manual / LLM-assisted)

1. Implement real tool logic in src/core/tools.mjs based on blueprint tool definitions.
2. Write prompt pack content (prompts/*.md) based on agent scope and security policy.
3. Write acceptance test cases based on acceptance.scenarios[].
4. See reference/stage_d_implementation_guide.md for patterns and best practices.

Stage D is performed manually by the developer or with LLM assistance. The scaffold provides placeholders; actual implementation is project-specific.

Stage E — Verify + Cleanup

1. Run verify command to execute acceptance scenarios.
2. Generate verification evidence (JSON + Markdown).
3. Review and fix any failures.
4. Update docs/runbook if needed.
5. Ensure registry entry is correct.
6. Delete the temporary workdir via finish --apply.

Verification

• Stage A docs exist in the temporary workdir and not in the repo.
• Stage B blueprint validates (validate-blueprint).
• Stage C scaffold created expected files without overwriting.
• Stage E verification completed and evidence files generated.
• Registry entry created/updated as specified.

Boundaries

• MUST NOT write to the repo during Stage A.
• MUST NOT proceed from Stage A→B or Stage B→C without explicit user approval.
• MUST NOT commit secrets to the repo.
• MUST NOT change fixed API route names (run, health).
• SHOULD NOT skip verification evidence generation.

Helper Tool Commands

Use the dependency-free helper script:

.ai/skills/workflows/agent/agent-builder/scripts/agent-builder.mjs

Reference Documents

Output Expectations

When executing this skill, always produce:

1. Stage A: Interview notes and integration decision (in workdir, not repo).
2. Stage B: Validated blueprint JSON (in workdir).
3. Stage C: Scaffold plan (file list) before applying, then apply summary.
4. Stage D: Implemented tools, prompt pack, and test cases (in repo).
5. Stage E: Verification evidence (JSON + Markdown) and delivery summary.

Final Delivery Summary Template

## Agent Delivery Summary

### Agent
- ID: {{agent_id}}
- Name: {{agent_name}}
- Module Path: {{agent_module_path}}

### Interfaces
| Type | Entrypoint | Response Mode |
|------|------------|---------------|
| http | node src/adapters/http/server.mjs | blocking |
| http (SSE) | POST /run/stream | streaming |
| http (WS) | ws://.../ws | streaming |
| ... | ... | ... |

### Verification
- Scenarios: {{passed}}/{{total}} passed
- Evidence: agents/{{agent_id}}/doc/verification-report.md

### Next Steps
- [ ] Configure environment variables (see .env.example)
- [ ] Deploy to target environment
- [ ] Enable kill switch (AGENT_ENABLED=true)