Multi-agent
Architecture where multiple specialized AI agents collaborate or relay on the same task or project.
Definition
A multi-agent system uses multiple agent instances, each with a defined role or specialization, to handle a complex problem. Agents can work in parallel (independent tmux sessions) or in sequence (handoff via context files). Each agent has its own context window and shares no memory state with others.
postcursors perspective
Multi-agent in daily dev is more prosaic than theory suggests. In practice, it's often a tmux split with two independent agent sessions — one on backend, the other on tests. No sophisticated orchestrator, no inter-agent communication. The context file (AGENTS.md) is the only link between sessions.
In practice
"Multi-agent workflow" means using multiple parallel agent sessions on the same project, each focused on a precise scope. Handoff between sessions happens via verifiable files (refactoring plan, TODO list).
Common misconceptions
- ✗ Confusing multi-agent with multi-model — using Claude then DeepSeek on the same task isn't multi-agent
- ✗ Expecting automatic inter-agent communication — in practice the developer is the orchestrator
In daily use, multi-agent in production looks like this: two tmux splits, two independent sessions, a REFACTOR_PLAN.md file as the link. The human decides who does what. The agents execute without talking to each other.