Agent orchestration is the layer that coordinates one or more AI agents. The orchestrator assigns tasks, passes results from one step to the next, keeps an eye on limits and decides when to stop or bring in a person. The orchestrator can be code, an automation platform or an agent that hands out subtasks to other agents.
Orchestration usually covers:
- Routing: which agent or tool picks up the task
- Step order and parallel branches
- Passing context between agents
- Step and time limits, error handling
- Checkpoints where a person approves the decision
In Pachca's breakdown of its own agent architecture, orchestration is simple: a loop of model calls with tools, a 15-step limit, a 5-minute timeout and a rule against sending more than two messages in a row without a reply from the user. The agent coordinates with people in the thread: it reports back there and can @mention the right teammate, who sees the whole history thanks to cross-chat threads.
Learn more in the Help center
Read the article →For developers
Docs on dev.pachca.com →