Official References: Sub-agents · Skills · Automation
Intermediate delivery is mostly workflow discipline
At week-scale, most failures are not model quality issues. They are handoff and sequencing issues.
Use a repeatable lane structure:
- planning lane
- implementation lane
- verification lane
Step 1: freeze scope before execution
Write down:
- in-scope modules
- out-of-scope modules
- acceptance checks
Step 2: run subagents where independence exists
Delegate noisy exploration and keep the main thread clean.
Step 3: standardize handoff fields
Minimum handoff:
- goal status
- evidence
- changed scope
- residual risk
- next owner
Step 4: end each day with evidence snapshot
Do not rely on memory between sessions. Capture command output and unresolved items daily.
Intermediate anti-patterns
Subagents for tightly coupled work
Parallelism adds overhead when dependencies are sequential.
Daily progress without verification
Progress reports become narrative, not evidence.
Scope drift accepted as normal
Week-scale work collapses when boundaries blur.
Quick checklist
- scope frozen and visible
- lane ownership explicit
- daily evidence snapshots captured
- final verifier lane run before merge
Intermediate reliability comes from repeatable rhythm, not heroic effort.