| name | multi-agent-orchestrator |
| description | Orchestrates the repository's custom SDLC agents through a staged workflow. Use when coordinating requirements, architecture, security review, API design, implementation, testing, documentation or final critique across multiple agents. Trigger phrases include orchestrate agents, multi-agent workflow, run the SDLC agents, coordinate backend frontend security testing, or critique the created agents. |
Multi-Agent Orchestrator
Use this skill when one request spans several roles and you want a deliberate sequence instead of ad hoc handoffs.
Workflow
- Start with
requirements-engineer when the request is still ambiguous or missing acceptance criteria.
- Use
platform-architect when ownership, system boundaries or trade-offs are unclear.
- Use
security-specialist before implementation for auth, data exposure or threat-surface review.
- Use
api-architect when endpoint contracts or integration schemas need to be defined before coding.
- Use
backend-specialist and frontend-specialist only after the expected behavior is concrete enough to implement.
- Use
test-specialist to add or validate meaningful checks for the changed behavior.
- Use
documentation-steward to update durable docs, ADRs, runbooks or release notes.
- Finish with
quality-engineer to critique the final output, missing validation and merge readiness.
Handoff Contract
Each stage should hand the next stage a compact artifact instead of raw chat history:
- Requirements: problem statement, user roles, acceptance criteria, open questions.
- Architecture: decision summary, ownership split, constraints, alternatives.
- Security: findings, remediation gates, unresolved risks.
- API design: endpoint or event contract, schemas, compatibility notes.
- Implementation: touched files, behavior changes, commands run, residual risk.
- Testing: coverage added, commands run, gaps that remain.
- Documentation: files updated and operational implications.
- Quality review: severity-ordered findings and missing evidence.
Operating Rules
- Do not skip straight to implementation when behavior, ownership or security posture is still unclear.
- Keep the number of active agents low. Prefer one stage at a time unless the work can safely branch into backend and frontend in parallel.
- When the repository contains an implementation surface for the requested change, the implementation stage must make code edits, run focused validation, and return concrete file changes instead of stopping at a plan.
- Only fall back to implementation planning when the workspace does not contain the relevant application code, runtime, or executable validation surface for the requested change.
- Require validation evidence before considering the workflow complete.
- If a stage discovers missing upstream inputs, route back one step instead of guessing.
Implementation Expectation
backend-specialist, frontend-specialist, and test-specialist are execution stages, not planning-only stages.
- Their default outcome is changed code, changed tests, and verification evidence.
- If a repo is documentation-only or otherwise lacks the implementation surface needed for the task, explicitly say that and produce the smallest useful implementation plan or contract artifact instead of pretending code was delivered.
Demo Invocation
Use the sample prompt in .github/prompts/run-multi-agent-demo.prompt.md together with docs/demo/sample-feature-request.md to test this workflow in chat.