agent-architecture
Installation
SKILL.md
Agent Architecture
Design multi-agent systems where coordination does not become the failure mode.
Process
- Justify the second agent. Most tasks that look multi-agent are one agent with better tools. Split only when roles need genuinely different context or run concurrently.
- Decompose by context boundary, not by job title. An agent exists to hold a distinct working set, not to mirror an org chart.
- Define the contract between agents — the exact shape of what is passed, and what is guaranteed about it. Free-text handoff is where multi-agent systems lose information.
- Decide where state lives. Shared mutable state between agents needs the same care as between threads.
- Isolate failure. One agent failing or looping must not stall the others; set timeouts and a supervisor that can terminate.
- Make the trace reconstructable. When output is wrong, you need to see which agent decided what, with what input.
- Bound the total cost. Agents that can spawn agents need a hard ceiling.
Deliverables
- Role decomposition with the context boundary justifying each agent
- Message contracts between agents
- Failure isolation, timeouts and supervision model
- Trace format and cost ceiling