Software Architect Crew
A 3-agent crew (PM → Coder → Reviewer) that takes a product prompt to reviewed code. Mermaid.js flowchart shows which agent currently holds the token; each agent's reasoning streams live over SSE.
A demonstration of multi-agent orchestration: three specialised agents collaborate to go from a natural-language product prompt to reviewed code in Python or TypeScript.
The crew
| Agent | Role | Model |
|---|---|---|
| Product Manager | Interprets the prompt, writes structured requirements | Claude Haiku 4.5 |
| Coder | Generates implementation from the requirements | Claude Sonnet 4.5 |
| Reviewer | Finds bugs and decides REVISE vs APPROVE | Claude Haiku 4.5 |
The feedback loop between Reviewer and Coder mirrors a real engineering team — the crew self-corrects until the Reviewer approves (capped at a sensible max cycles).
Visual proof: live flowchart
A Mermaid.js flowchart rendered inside the Astro page highlights which agent currently “holds the token” — updating in real time as the SSE stream emits state events. The chart makes the orchestration legible to non-technical visitors.
Tech choices
- Custom orchestrator as an async generator — no third-party multi-agent framework. The graph is small enough that explicit code is clearer than a runtime abstraction.
- Vercel AI SDK (
ai) for all model calls. Quality work (code generation) routes to Claude Sonnet; fast structured work (requirements, review) routes to Claude Haiku. - Server-Sent Events for the live trace — every thought, every state transition, every artifact streams to the browser as the run progresses.
- Mermaid.js renders the flowchart client-side, fed by the live
stateevents.
Why this matters
This project draws on my experience as a Scrum Master at the Bank of England — translating the human ceremonies of sprint planning, refinement, and review into an automated multi-agent workflow you can watch run end-to-end.