Multi-agent orchestration

At its core, AIgency is a dependency-aware, concurrent scheduler. It decides which of a run's tasks can start now, respects the dependency graph between them, and keeps as many agents working in parallel as your concurrency settings allow.

Concurrency

A run has a global concurrency ceiling (how many tasks may run at once across all agents) and per-agent concurrency ceilings (how many tasks a single agent may run at once). The scheduler tracks peak observed concurrency for a run, so you can see how parallel it actually got, not just how parallel it was allowed to be.

Failure policy

Each run has a failure policy:

• isolate — a failed task doesn't stop the rest of the run; independent tasks keep going, and only tasks that depended on the failed one are affected.
• cancel_all — a single task failure halts the whole run.

The activity feed

While a run is active, AIgency maintains a most-recent-first activity/event feed (capped at 200 events) so you can watch what's happening across every task without switching between individual agent conversations. For the full historical record after a run finishes, see Run history.

Related

• Execution covers what actually happens to an individual task as it runs, including isolated git worktrees.
• Agent assignment covers which agent a task is scheduled on.