Set up a team of AI agents. Keep people in charge of the decisions.

1. 01

   Settings → Compute → AI orchestration

   Start with the default workflow

   New organizations start with Design, Risk Assessment, Impact Analysis, Code Generation, Test Generation, Code Review and AI Test & Iteration. Keep these defaults for a first run, or open a role to add instructions and make the workflow match your team.

2. 02

   Settings → Compute → AI orchestration · Project settings → AI providers

   Confirm a model and credential path

   Before starting, confirm the default model has a usable organization or project provider key, or assign a named credential profile to a role. ReasonOS stores a named key sealed in the control plane; its raw value is never shown again or passed to a branch server or agent prompt. Missing credentials prevent a start; provider quota failures are reported during execution.

3. 03

   Customize a role

   Give each role the right team and authority

   Choose a primary agent, add independent reviewers when a second point of view helps, and set each agent’s engine, model and credential policy. Set a named person or a role manager as final authority, or let AI carry its recommendation forward. Role authority never approves or merges code.

4. 04

   Project settings → AI orchestration

   Tailor a project without changing every project

   A project inherits the organization workflow until you override a standard role or add a project-specific one—for example, a security architecture review. The project page makes the inherited policy, enabled agents and human decision gates visible before a run starts.

5. 05

   Task detail → Run AI workflow

   Run the workflow from an active task

   Open a task in a running branch and start the configured workflow. ReasonOS snapshots the effective roles and agent assignments for that run. Independent, read-only work can run in parallel; code, test-generation and AI-test work are serialized so two writers do not race in the same branch.

6. 06

   Task detail → Workflow evidence

   Review the evidence and make human decisions

   The task records agent responses, verdicts, policy transitions and any recorded default-credential fallback. When a human-authority role is ready—or the system needs help—you review that evidence, then choose Approve and continue or Stop workflow.

7. 07

   Customize AI Test & Iteration

   Use AI Test & Iteration to repair bounded failures

   Set zero to turn automatic repair off, or choose a limit from one to ten repairs. An actionable failure can enter a Design → Code → AI Test loop. A passing test completes the run; a flaky, nonconvergent or exhausted run stops with its evidence for a human decision.

8. 08

   Task detail → Rate agent · Project settings → Activity and evaluation

   Learn from observed outcomes

   Score a completed attempt from 0 to 100 based on the actual task result. The project report shows active, completed and failed attempts and each configured agent’s average score, so comparisons are grounded in your work instead of a generic benchmark.

Start conservatively: leave final authority with a person, use a verified default model path for an initial run, and add independent reviewers where a second opinion has real value. Your existing branch, review and merge rules remain in force throughout.