Parallel Research Synthesis
fork + agent: Explore: Three Haiku forks in parallel, Opus synthesizes. ~30s. A production architecture.
Parallel Research Synthesis pairs context isolation with model-routing to spawn multiple cheap researchers in parallel and orchestrate their results in a single costly synthesis pass.
How It Works
The pattern combines two Claude Code primitives: context: fork (isolated execution per invocation) and agent field (task-specific model selection). A meta-skill invokes multiple research skills in parallel; each runs in its own fork with agent: Explore (routes to Haiku). Each fork produces focused findings at low cost. The parent orchestrator, running on Opus, collects results and synthesizes them into a coherent answer.
The routing happens at invocation time: each skill specifies which agent to use, and Claude Code ensures it runs under that agent's model and tool set. Fork isolation prevents context leakage between parallel runs.
Test Results
Runtime test (documented in experiments/claude-features/2.1.0/WORKFLOW-IDEAS.md): three Haiku forks invoked in parallel, Opus orchestrator synthesized results. Total runtime ~30 seconds. Test confirms both the parallel execution and the model-routing isolation work as expected.
Why It Matters
Research and synthesis are computationally asymmetric: finding candidates is cheap and embarrassingly parallel; combining findings into coherent output requires costly reasoning. This pattern makes that asymmetry structural. Haiku can handle divergent, exploratory search tasks (broad web scanning, fact-gathering, candidate enumeration). Opus can focus only on the cognitive work that justifies its price.
The cost savings are significant: a three-way parallel search on Haiku costs roughly one-ninth of a single Opus run, and the results are synthesized (not compromised) because synthesis is the expensive task.
Mechanics and Constraints
Each forked skill must specify context: fork in its frontmatter and declare an agent field pointing to a custom agent or agent type (e.g., Explore). The parent skill or workflow does not fork; it blocks until all parallel research skills return, then synthesizes.
Fork isolation means no shared session state across researchers. If researchers need common context (e.g., a corpus to search), that context must be passed as explicit input to each skill.
- file2.1.0/WORKFLOW-IDEAS.md
workflow_bd9f4a1c5d8da796ea46ec9ded255199b87705613b1e2fd064d57fa75a6b679d2856ceafad6b1daa8f982493871b6dd7a4a96cf4236a17e1c039856a134d9d30f78f0d4e195e65d7aaeba8273923501138054d0b9dffbafe15caaa5123bf407b9bccad8295b079c79f4834cd448d204Signed with an ed25519 key held off the repo. Anyone can verify against the published public key; nobody without the secret key can forge it. Click verify: it recomputes the signature in your browser. The signature proves integrity and authorship of this exact content — not a third-party timestamp or that the underlying claim is objectively true. signedAt is when the @f3/attest pipeline ran, not when the work happened; the evidence refs carry the source dates.