Forked parallel agents can coordinate work through shared task state and metadata updates, enabling divide-and-conquer exploration.

Test Setup

Two agents (FORK-EXPLORER-A and FORK-EXPLORER-B) were spawned in parallel, each assigned a separate investigation domain: task system architecture and session ID discovery. Both agents were instructed to create tasks, research independently, and persist findings via task metadata updates. An orchestrator spawned and monitored both agents while maintaining its own task list.

What the Test Found

Both agents successfully created tasks visible to all other agents in the shared TaskList. FORK-EXPLORER-A created task #21; FORK-EXPLORER-B created task #22. Each agent updated its task metadata with research findings, including architecture details and session ID discovery methods. All agents operated on the same local task store at ~/.claude/tasks/{session}/, confirming shared state.

The test discovered three operational session ID vectors: hooks (JSON stdin), the sessions-index.json file, and skill string substitution. Session IDs were confirmed absent from environment variables.

The Pattern

The test demonstrates a repeatable workflow: create "exploration" tasks for distinct hypotheses, spawn parallel agents each claiming one task, have agents research independently and update task metadata with findings, then have the orchestrator synthesize results from all metadata. Findings persist even if an agent process terminates early.

Limitations

The test validated task visibility and metadata coordination but did not measure convergence time, handle agent failures mid-research, verify metadata consistency under concurrent updates, or test at scale beyond two agents. Whether this pattern remains reliable under high coordination load or with longer-lived forked processes remains open.