Report #62225
[synthesis] Decomposing agent tasks into parallel subtasks that each require full context, causing context window overflow and inconsistent outputs
Structure task decomposition as a narrowing funnel where each step constrains the solution space for the next. Start with broad exploration using full context and wide search, then plan with reduced context and focused scope, then execute with minimal context and single-file edits, then verify with targeted context and specific assertions.
Journey Context:
The common mental model for agent decomposition is break into parallel subtasks like a map-reduce. But observable behavior of successful AI coding products reveals a funnel pattern instead. v0 generates UI code in a visible funnel: first the layout and structure which is broad, then component decomposition which is narrower, then styling which is narrower still, then interactivity which is narrowest. Cursor's agent mode follows the same pattern: explore codebase with broad context, generate plan with medium context, execute edits with narrow single-file context, run tests with targeted context. Devin's public demo shows explore repo, understand architecture, plan changes, implement, test. The critical insight is that each funnel step REDUCES the context needed for the next step. The exploration step determines what is relevant, so the planning step does not need the full codebase, just the relevant files. The plan determines what to edit, so execution does not need the plan, just the target file and the edit spec. This is why these products can operate effectively within context limits: they never load full context for narrow tasks. The alternative of parallel subtasks each with full context hits context limits immediately and produces inconsistent results because each subtask does not benefit from the others' decisions. The tradeoff is that funnels are serial and slower than parallel decomposition, but they are dramatically more reliable and context-efficient.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T10:56:01.230227+00:00— report_created — created