flovencatalasan Adaptive Reasoning
Match reasoning effort to the work. Use the lowest level likely to complete the task reliably.
Choose a role
- Scout — GPT-5.6 Sol Low: Answer a narrow, read-only question such as locating files, tracing one code path, finding relevant tests, or checking a focused fact. Retain enough judgment to discover useful adjacent context.
- Worker — GPT-5.6 Sol Medium: Implement a scoped change, run checks, diagnose a routine defect, or complete supporting work with clear boundaries.
- Smart worker — GPT-5.6 Sol High: Handle difficult implementation, cross-cutting reasoning, consequential ambiguity, or a task where a lower effort is likely to miss important constraints.
Handle trivial questions and one-step edits directly without delegation.
Execute sequentially
- Classify the next substantive unit of work before delegating.
- Spawn exactly one worker using
gpt-5.6-sol, the selectedreasoning_effort, andfork_turns: "none". Give it a self-contained objective, relevant paths or sources, constraints, success criteria, and required output. Instruct it not to delegate. - Wait for that worker to finish before spawning another. Never run workers concurrently.
- For mixed work, run roles sequentially only when needed—for example, a Low scout followed by a Medium or High worker using the scout's findings.
- Integrate and verify the result in the primary agent. Keep approvals and consequential decisions with the user.
Escalate reasoning only when task complexity or a concrete failed attempt justifies it. Do not repeat completed work at a higher effort merely for reassurance. If the requested model or effort is unavailable, use the closest available equivalent and state the substitution.