biggora Delegate Bounded Work
Keep orchestration small, explicit, and evidence-driven. Retain end-to-end responsibility as the orchestrator.
Choose the Execution Mode
- Default to single-agent execution for a small coherent edit, a local investigation, or work whose context and verification already fit one agent. Avoid delegation when coordination cost exceeds its measurable benefit.
- Choose Mode 1: Bounded Delegation when one separable deliverable benefits from specialist knowledge, parallel discovery, context isolation, or a distinct tool surface.
- Choose Mode 2: Independent Fresh-Context Review when an existing artifact needs unbiased acceptance verification.
- State the chosen mode, the reason it adds value, assumptions, task boundary, and success evidence. Do not create a swarm to compensate for unclear scope.
Mode 1: Bounded Delegation
Delegate one atomic deliverable to one appropriate specialist. Keep dependent or cross-cutting integration with the orchestrator unless another boundary is independently verifiable.
Before dispatch, write a contract containing:
- Role: one specialized responsibility.
- Context: only repository and product facts needed for this deliverable.
- Task: exact actions, locations, and behavior.
- Constraints, tools, and authority: required skills/tools, quality rules, allowed mutations, and actions that require approval.
- Input: exact files, data, artifacts, or prior verified outputs.
- Expected output and acceptance criteria: deliverable format plus measurable proof.
- Example: one representative correct result or stable reference.
- Stop condition: completion boundary and conditions that require returning control rather than expanding scope.
Require the delegate to return the artifact or precise artifact location, a concise result summary, verification evidence, assumptions, and limitations. Do not request a full reasoning trace.
Inspect the result against every contract field. Grade it PASS, FAIL, or PARTIAL. Accept only PASS; for FAIL or PARTIAL, re-delegate with the failed criterion, observed evidence, and correction boundary instead of silently patching the delegate's work. Escalate a repeated identical blocker rather than adding agents or broadening authority.
Mode 2: Independent Fresh-Context Review
Start a fresh reviewer only after the artifact exists. Give the reviewer only:
- the artifact or diff;
- the approved specification and acceptance criteria;
- commands or steps required to run it;
- verifier tools and environment facts needed for black-box checks.
Do not provide implementation history, suspected defects, expected answers, prior conclusions, or the implementer's reasoning. Ask the reviewer to test acceptance criteria as a user or external verifier and flag only correctness or requirements failures. Exclude style preferences, speculative improvements, and unrelated refactors.
Route web journeys through test-web-ui, mobile journeys through test-mobile-app, and visual conformance through ui-audit when those skills are available and applicable. Keep the review read-only unless the user separately authorizes remediation.
If a fresh agent is unavailable, prepare the same review brief, run the black-box checks locally, label the result Independence: MISSING, and require human or external review before claiming independent sign-off.
Grade the review PASS, FAIL, or PARTIAL against the approved criteria. Re-delegate remediation through a bounded contract; never leak the expected fix into the next independent review.
Assemble Verified Work
Integrate only accepted artifacts, preserve unrelated changes, and rerun the relevant end-to-end checks after integration. Report each contract or review, its grade, evidence, retries, and unresolved limitations.
Read delegation and review briefs before dispatching work or independent review.