This is preview documentation. The official release is not yet available.
Team and member fields
A Researcher should provide sources and findings; a Reviewer should identify evidence gaps. The Leader resolves disagreements, states unfinished work and submits one final result. Team roles do not replace the Agent’s own model, tool or resource configuration.
Coordination limits
Outside-roster delegation is unrelated to the External Agent runtime mode. An External Agent included in the roster is still a Team member.
The current creation form writes explicit values: concurrency 32, fanout 8, hops 8, child depth 8 and child issues 64. These creation values differ from zero-value handling above. Inspect the saved policy rather than interpreting every zero as unlimited or disabled.
A small-team policy fragment:
expectedVersion checks. Concurrency, depth and roster limits are not a hard cap on external model charges; manage usage through the actual providers and account policies.
Mixing runtime modes
Leaders need coordination capabilities: delegation, reading results and node completion/failure. Members need their assigned role’s capabilities. Chat availability alone does not qualify an Agent as Leader.
Runtime policy
runtimeBindingPolicy contains ordered candidates, selectionMode, fallbackMode and optional retryPolicy. Each candidate’s binding selects the backend; requiredCapabilities and securityConstraints constrain selection. Use valid binding identities from the console/API, not process IDs or arbitrary URLs.
Selection follows node override, member override and Agent policy layers. Explicit fresh fallback creates new execution context, so retain evidence in Issues, comments and Artifacts. Verify one candidate before adding fallbacks. Scaling and fallback do not resolve shared-file conflicts automatically.
Next: collaboration guide and execution model.