Infinite Game OS

Definition

Researcher runs parallel sub-agent research on any topic. Up to 4 independent workers spawn simultaneously, each assigned a specific research angle with relevant context and a clear output format. Workers do not communicate with each other; only their final outputs return. The skill aggregates worker outputs into a single structured report (summary, findings by angle, recommended actions or decision points, sources). Research is intelligence-gathering only. Workers do not implement, do not write to canonical files, do not trigger external actions. Every report returns to the operator for review and decision.

Parallel workers, synthesized report

Phase 1 clarifies angles before spawning. If the request did not specify, the skill proposes angles (typical patterns: what exists vs what is missing; best practices vs common failure modes; compatible vs incompatible patterns; short-term vs long-horizon considerations). Phase 2 spawns up to four workers in parallel, each with a specific angle, context and output format. Workers are independent; sub-agent depth caps at 1.

Phase 3 synthesizes worker outputs into a single structured report (summary, findings by angle, recommended actions for operator review, sources). Phase 4 presents the report. The operator decides what moves forward; the skill never auto-implements findings. External web research via WebSearch and WebFetch is allowed; external writes are not. Every report returns to the operator. That boundary is non-negotiable.

Use Cases

Pre-plan deep dive across multiple angles

Before committing to a major plan design, the operator dispatches Researcher on the topic. Four parallel workers investigate problem landscape, existing patterns, risk surface and adoption examples. The synthesized report feeds the plan design with grounded context rather than first-principle guessing.

Tech-watch comparative scan across multiple repos

Watched repos accumulate over time. Researcher spawns workers per repo to surface new patterns, hook types and skills since last review. The report shows what changed across the watched set in one pass rather than serial reviews.

Architectural decision spanning multiple protocols

A decision touches multiple systems. Workers cover existing protocols for conflicts, relevant references, external precedents and implementation complexity. The report surfaces dependencies the operator would have missed in linear research, plus a Decision Points section for what needs explicit choice.

FAQ

Why parallel workers instead of one deep agent?

Parallel workers cover more ground per unit of time and surface convergence that one agent would miss. When two workers independently flag the same finding, that overlap is itself a result. A single deep agent produces a linear report; parallel workers produce a triangulated one.

Can workers act on findings?

No. Workers gather and report. They do not implement, do not write to canonical files, do not trigger external actions. Every report returns to the operator for review and decision. That boundary is non-negotiable.

What is the maximum sub-agent depth?

Sub-agent depth is capped at 1. Workers do not spawn nested sub-agents within themselves. This prevents recursive cost spirals and keeps the research surface area observable.