Guide · Agent orchestration
Route Agent Work by Difficulty, Evidence, and Consequence
Create clear escalation rules so routine tasks stay simple and difficult decisions reach a worker equipped to handle them.
A routing rule needs more detail than "use the strongest model for hard tasks." The system has to recognize what makes a task difficult and what evidence justifies escalation. Start with a few observable conditions, then review the cases your initial rules get wrong.
Define a routine path
For a hypothetical support-draft assistant, a routine request might have a known issue type, a matching current help article, and enough account information to prepare a draft. The worker can cite the article and return the draft for the usual review.
An unfamiliar error, conflicting documentation, or missing account detail belongs on a different path. Those are conditions the application can record. A model's confidence can be another input, but it should not be the only signal used to decide whether a result is dependable.
Keep the routine assignment narrow. If the worker is allowed to redefine the problem freely, it may expand a simple draft into an investigation that the routing rule was never designed to manage.
Escalate with a reason and a packet
When the task needs a different worker, include the original request, the work already attempted, relevant evidence, and the specific unresolved question. A more capable model should not have to repeat all the discovery just to learn why it was called.
You can evaluate GPT-6 Astra or Claude Fable 5.1 for the difficult branch using the same escalation cases. Choose from observed results on that role. Do not assume that a higher-priced configuration will resolve missing inputs or grant access to a tool the system does not have.
Use escalation reasons you can count later, such as contradictory sources, failed reproduction, or a requirement outside the routine workflow. That makes it possible to see whether the difficult branch is handling real exceptions or compensating for an unclear initial brief.
Separate difficulty from permission
A task can be easy to reason about and still have a significant consequence. Sending a prepared message may be technically simple, while changing customer data may require an explicit approval. Route those actions according to the permission rules even when the reasoning is routine.
Likewise, an advanced model does not need every tool. Give the escalated worker the access needed for its assignment and preserve the existing limits on publishing, deleting, or spending. A role change should be visible in the task record.
Review the routing mistakes
Collect examples of unnecessary escalation, missed escalation, and tasks that bounced between workers. Ask whether the trigger was unclear, the evidence was missing, or the receiving assignment was too broad. Change one part of the routing at a time and replay those cases.
Keep a human fallback for situations the workflow cannot resolve responsibly. A useful fallback includes what is known, what remains uncertain, and the decision needed to continue. It should help someone finish the job without reading every worker's conversation.