Agent orchestration

AI systemsDefined term⦿ live instrument below

Orchestration is everything around the loop: decomposing a goal into steps, routing them to the right model or sub-agent, carrying state between calls, bounding cost and time, and deciding what happens on failure. Workflow automations are orchestration with the model as one node among many.

The engineering trade-off is autonomy versus determinism. Model-driven control adapts but compounds error; explicit orchestration (state machines, DAGs, workflow engines) is auditable and testable. Production systems overwhelmingly push control flow out of the model and reserve the model for judgment.

Step 1 / 8
  1. USER GOAL

    Goal: “Which of our 3 suppliers delivered fastest last month, on average?”

Context window40 / 800 tokens

Every observation is appended — the loop pays for its own history on every subsequent model call.

Per-step reliability95%

Task success ≈ p^n: at 95% per step, a 1-step run completes 95% of the time — and a 12-step run only 54%.

Live instrument — see Agent orchestration behave.