What it does

Generative orchestration lets the model plan a response by choosing which topics, actions and knowledge sources to use, and chaining them, instead of you wiring every path by hand. You describe capabilities; the model decides how to combine them.

Key facts

• Turned on per agent. With it off, the agent runs classic topic-matching only.
• The model selects tools and knowledge based on their names and descriptions — those descriptions are effectively prompts.
• It can gather missing inputs across multiple tools in one turn, filling parameters from the conversation.
• Instructions at the agent level steer orchestration behaviour and boundaries.

When to use / skip

Use it when an agent has several tools and knowledge sources and rigidly scripting every combination would be unmaintainable. Skip it for narrow, deterministic agents where you want tight control and predictable, repeatable paths.

Configuration decisions

• Orchestration on or off, per agent.
• How you name and describe every tool, topic and knowledge source so selection is reliable.
• Which deterministic topics to keep as guardrails around the model's freedom.

Gotchas

• Poor tool descriptions are the number-one cause of the model calling the wrong thing or nothing.
• Behaviour is less predictable than scripted flows — invest in test sets, not one-off manual checks.

Consultant notes

• This is the strategic direction, but it demands discipline in descriptions and testing. Treat naming as a design activity.
• Non-determinism has governance implications — evaluate before enabling on regulated processes.
• Expect higher message consumption from multi-step reasoning; model capacity accordingly.

---

Revisit as orchestration capabilities mature each wave.