Agent Orchestration

When more than one agent, model or tool has to cooperate on a task, something has to decide the order, carry the state and clean up after failures. Agent orchestration software provides that layer: routing between specialised agents, planning and task decomposition, shared and persistent memory, tool and function registries, concurrency limits, retries with backoff, timeouts, checkpointing so a long run can resume, and tracing that shows every step, prompt and tool call after the fact.

Platform and infrastructure teams are the buyers, usually once a prototype built on a single loop has started failing in ways nobody can explain. Real differences: whether workflows are expressed as a graph, a state machine or a conversation; durability across restarts; depth of the tracing; model and provider neutrality; self-hosting; token and cost accounting per run; and support for approval gates before an irreversible action.

Judge one on a failed run rather than a successful demo — how quickly the cause is found, and whether the run can be replayed from the failing step. Non-determinism makes testing awkward, costs escalate quietly, and proprietary abstractions are expensive to leave. Licensing tends to be open source with a paid hosted control plane, priced per run or per seat.