Claude Code Parallel Agents

Definition

How to run Claude Code work in parallel (captured from code.claude.com/docs/en/agents, 2026-07-31, HTTP 200). Four distinct approaches, each differing in who coordinates and who the workers report to.

Key Points

• Four ways to parallelize, not one feature: subagents (delegated workers inside one session, returning summaries), agent view (claude agents, one screen to dispatch/monitor background sessions, Research preview), agent teams (coordinated sessions with shared task list + inter-agent messaging, lead-managed, Experimental/disabled by default), and dynamic workflows (a script running many subagents and cross-checking results).
• Selection driven by 3 questions: who coordinates, whether workers must communicate, and whether they touch the same files. Subagents report to the spawning conversation; agent-view sessions report only to you; teammates message each other.
• Worktrees are the file-conflict answer — parallel sessions must never edit the same files; agent view moves each dispatched session into its own worktree automatically; agent teams do not isolate, so partition files yourself. /batch packages 5–30 worktree-isolated subagents, each opening a PR.
• Monitoring differs per approach — claude agents opens agent view; named background subagents appear in the @-mention typeahead; /tasks lists background items; /workflows lists runs/phases/completed agents.
• Parallelism multiplies token cost — explicitly documented, with pointer to /docs/en/costs for usage and rate-limit details.
• Gauntlet-loop as userland orchestration — the RoboNuggets transcript shows an externalized prompting pattern that uses parallel worker+critic pairs and a high stopping bar, effectively packaging subagent fan-out + evaluator loops as one reusable prompt structure rather than as a built-in Claude Code feature.

Implications

Parallelism is a first-class token-spend multiplier — it must be budgeted against the token-usage-reduction levers, not treated as free. The subagent-summary pattern is the vault’s strongest context-isolation example: verbose output stays in the worker’s context window and only the summary returns. Worktree isolation is the safety precondition for parallel sessions to be viable. The gauntlet-loop example matters because it shows how these mechanics become a prompt-level workflow pattern that non-experts can copy, not just a product feature hidden in docs.

Open Questions

• Version-sensitive: as of v2.1.198 /agents no longer opens a panel (prints a notice); /subtask vs /fork depends on whether agent view is on; agent teams experimental/disabled by default; agent view is Research preview.

Related

claude-code-sessions — background sessions, /tasks, parallel session coordination token-usage-reduction — explicit “multiplies token usage” cost lever claude-code-system-prompt — subagents as specialists defined by system prompt + tool restrictions ai-agents-guide — general agent coordination context claude-code-subagents — the delegated-worker mechanism gauntlet-loop-prompting — prompt-level worker/critic loop built on parallel subagent orchestration

Sources

• raw/external/code-claude-com-agents-47001d98.md
• raw/prompts/articles/robonuggets-gauntlet-loop-youtube-transcript.md