Phase 5: Implement with Agent Swarm¶

Goal

Deploy coordinated agents to execute beads systematically.

This is where the planning pays off. With a well-QA'd bead graph, multiple agents can work in parallel without conflicts.

Prompt: Swarm Marching Orders¶

Give this to each agent you spin up.

Marching orders (give to every agent)

First read ALL of `AGENTS.md` and `README.md` super carefully and understand
ALL of both.

Then:

- investigate the codebase to understand architecture and invariants
- use beads as the source of truth for what to do next
- use `bv --robot-next` / `bv --robot-triage` to select impactful actionable beads
- claim work by marking beads in_progress and keep them updated as you go
- avoid communication purgatory; start shipping work
- respond to any coordination messages promptly (if using an agent mailbox)

Proceed meticulously on the next assigned beads. Use ultrathink where available.

Prompt: Reread AGENTS (Context Recovery)¶

When an agent's context gets compacted or you suspect drift, re-anchor it.

Re-anchor prompt

Reread `AGENTS.md` so it's still fresh in your mind. Use ultrathink.

Agent Coordination Tips¶

Multi-agent best practices

Use Agent Mail (if available) for inter-agent messaging to prevent merge conflicts
Claim before starting -- each agent should mark beads in_progress before working
Communicate blockers and discoveries, not just status updates
Group commits by bead/feature slice, not by time
Avoid communication purgatory -- if an agent is spending more time coordinating than coding, something is wrong

Typical Agent Allocation¶

For a medium-to-large project, Emanuel typically deploys:

Agent type	Count	Role
Claude Code (Opus)	5-6	Primary implementation
Codex	2-3	Parallel implementation
Gemini	1-2	Review duties

Scale up or down based on your bead graph size and budget.

Stop condition

All beads are implemented and marked complete.