Guide · Agent orchestration
Use Astra and Fable 5.1 in One Workflow Without Mixing Their State
Keep a shared task record and explicit handoffs when OpenAI and Claude agents work on the same project.
You can design a workflow that uses GPT-6 Astra for one assignment and Claude Fable 5.1 for another. The important boundary is between their application state. A useful handoff contains the task, artifacts, evidence, and open questions. It should not depend on one provider understanding the other's private conversation format.
Let your application own the task record
For a hypothetical feature build, keep one record with the requirement, source revision, active assignments, accepted changes, and review status. Store provider-specific conversation identifiers alongside the relevant worker, rather than using one provider's conversation as the only record of the project.
That separation gives the coordinator something stable to inspect after a restart or model switch. If a worker becomes unavailable, another can continue from the latest accepted artifact and the remaining task. You still need to explain the context, but you do not need to reconstruct it from a long transcript.
Exchange artifacts with a small summary
An Astra implementation worker might return a patch, the tests it ran, and the known limitations. A Fable reviewer can receive those artifacts along with the original requirement and a clear review brief. This is an example role assignment to evaluate, not a claim that either model always belongs in that role.
Pass the evidence the next worker needs. A summary that says "everything passed" is less useful than the commands, results, and revision they applied to. Keep full logs available by reference, while putting the important result directly in the handoff.
Keep provider details behind adapters
Use separate request builders for each provider. Let those adapters translate your application's assignment into supported messages, tools, and settings. Do not forward an OpenAI response object directly into a Claude request and expect its internal fields to carry the same meaning.
Treat hidden reasoning and provider-specific thinking blocks as implementation details of that conversation. Share a concise account of decisions and evidence instead. The next worker needs to know what was concluded and why it is supported, not inherit an incompatible internal representation.
Give one component authority to accept work
Two models should not independently decide that a shared feature is ready to release. Let the coordinator or an explicit application step accept the combined result after the required checks. Keep external actions such as publishing behind the existing approval rules.
If the reviewer asks for a change, attach the finding to the exact revision it reviewed. After the implementer updates the patch, invalidate any review result that no longer applies. Otherwise, an old approval can accidentally travel with new code.
Keep the second provider only if it helps
Run a small comparison against the same workflow using one provider. Measure accepted findings, repair time, elapsed time, and total cost. Cross-provider orchestration adds another integration to maintain. It earns its place when the completed work improves enough to justify that responsibility.