Multi-agent — when one isn't enough

Sometimes one agent gets confused juggling 12 tools. Splitting helps:

- Researcher — finds info
- Coder — writes code
- Critic — reviews + suggests fixes
- Coordinator — routes between them

Each agent has fewer tools, a clearer role, focused prompt. They communicate by handoff (passing structured messages) or shared state.

Warning: multi-agent costs MORE than single-agent. Only worth it when role separation actually clarifies things. Don't multi-agent for vibes.

Two common shapes:

1. Pipeline — researcher → coder → critic → coordinator. Linear. Easy to reason about. LangGraph default.

2. Network — coordinator routes any agent to any other. Dynamic. Messy but flexible. AutoGen's strength.

Start with pipeline. Network gets you 'fascinating' demos and unbounded loops. Pipeline gets you something that works.

If the task fits a pipeline, use it. If not, the task often isn't ready for multi-agent.

Researcher gathers info. Coder uses it to write code. They pass a typed message between them. No framework — just function calls.

This is the Lego brick of multi-agent. Add a third critic by the same pattern.

Network topologies can ping-pong forever ('agent A asks B, B asks A back...').

My rule: max 12 inter-agent messages. After that, force the coordinator to produce a final answer with whatever info it has.

Log every handoff. When the system is weird, the handoff log shows you the dance. Without the log, multi-agent is opaque.

Day 74 done.

- Multi-agent = specialisation
- Pipeline > network for first version
- Two-agent pipeline in 30 lines
- Cap turns; log handoffs

Tomorrow (Day 75): the agent frameworks worth knowing — LangGraph, AutoGen, CrewAI.