Guide · Agent orchestration
Using GPT-6 Astra as an Agent Coordinator
Give GPT-6 Astra clear delegation rules, concrete completion criteria, and a manageable view of the work it coordinates.
A coordinator has a different job from the workers it supervises. It needs to understand the request, assign useful pieces, notice missing evidence, and assemble a result. If you use GPT-6 Astra in that role, write instructions for those decisions instead of assuming a capable model will infer your preferred workflow.
Tell it when delegation is useful
OpenAI's Astra guidance calls out delegation, clarification, and verification as behaviors worth tuning. In your own system, specify when a separate worker is justified. Independent codebase research or a focused review can be good candidates; a small edit with all the context already available may not need another agent.
Give the coordinator a practical threshold: delegate when the task has a clear output and can proceed independently of useful work in the main task. Ask it to keep the assignment narrow enough that the returned result can be checked. This makes delegation a decision about the work rather than a requirement to keep several agents busy.
Keep the coordinator's state small
For a hypothetical reporting tool, track the request, active assignments, accepted findings, open questions, and final artifact. The coordinator does not need every research log in its main conversation. It does need a path back to the evidence behind each important conclusion.
Ask workers to return what changed and what remains unresolved. If a worker sends a long narrative, the coordinator may miss the one dependency that matters. A concise result with artifact references is easier to use than a transcript of every step.
Make ordinary autonomy explicit
Describe which routine decisions the coordinator can make within the brief. It might choose an internal file layout or retry a failed read without interrupting the user. Identify the decisions that change the user's commitment, such as publishing a result or expanding the agreed scope.
Keep those boundaries in the application as well as the prompt. A coordinator's explanation of its intentions is useful for review, but access to consequential tools should still follow the permissions you actually configured.
Ask for evidence proportional to the task
A tiny wording change and a database migration need different checks. Define the acceptance criteria before dispatching work, and ask the coordinator to stop broadening verification once those criteria are met unless a new concern appears.
For the reporting example, that might mean confirming the requested date range, reconciling totals against the supplied data, and opening the finished file. It need not include unrelated improvements to the reporting system.
Review the orchestration, not just the prose
After a pilot, inspect the assignments Astra made. Were they independent? Did the coordinator use the returned evidence? Did it notice a blocked worker? Keep examples of unnecessary delegation and missed handoffs so the next prompt revision addresses observed behavior. A polished final answer should be the end of a sound process you can inspect.