- Published on
Claude Code v2.1.302: Subagents Now Fork by Default β Parallel Agent Workflows Evolve
Anyone who has run multiple AI agents simultaneously knows the frustration: outputs from one agent bleed into another''s context, threads get tangled, and token counts balloon. Claude Code v2.1.302 changes the default behavior that caused all of this.
What "Fork by Default" Means
In software, a fork copies an existing process and spins up an independent child. Claude Code''s fork works the same way.
Before v2.1.302, subagents either started fresh (knowing nothing about the parent conversation) or ran inline, piling all their tool output directly into the parent context. Large file reads or search results would flood the parent conversation.
From v2.1.302 onward, subagents fork by default. They inherit the entire parent conversation context at launch, but the tool output they generate during their run never flows back into the parent. Only the final result does.
What Actually Changes in Practice
Picture a manager sending a team member to research a topic. The old model was like having the team member''s laptop screen mirrored live on the manager''s screen. The forked model is the team member working in a separate room and delivering only the finished report.
Core differences:
- The parent context stays clean throughout the subagent''s run
- Intermediate steps (file reads, API calls, web searches) don''t accumulate in the parent window
- Multiple subagents can run in parallel without context collisions
- Only the final result is returned to the parent agent
Parallel Agent Workflow Example
Here is what this looks like in practice:
Research + coding simultaneously
Parent agent:
βββ Subagent A: "Research competitor API docs and summarize"
βββ Subagent B: "Analyze the authentication module in our codebase"
Both A and B read files and run searches, but none of that intermediate noise accumulates in the parent window. When both finish, only their summaries arrive.
Parallel code review Splitting a large pull request across multiple subagents β each reviewing a different section simultaneously β is now practical. Each agent works independently; the parent receives a consolidated report.
The Broader Direction in Claude Code
Fork-by-default is not an isolated change. It is part of Claude Code''s shift from a single-session coding assistant to a genuine multi-agent orchestration platform.
The claude_code.safety_event OpenTelemetry event introduced in v2.1.310 (October 23) fits the same picture. It fires each time a tool call is stopped or flagged by a permission rule, a safety check, a hook, or the sandbox β giving teams audit visibility across large agent pipelines.
Practical Ideas for Education and Work
- Curriculum development: research multiple subject areas in parallel and merge the findings into one report
- Multilingual content: route a single draft to separate subagents for each language translation, collect results
- Student code review: send submitted code to multiple subagents simultaneously, generate batch feedback
Tips
- Explicit
/agent forkis available but rarely needed now that forking is the default - Write subagent prompts clearly β since they inherit parent context, you can skip background re-explanation, but the task description itself must be precise
- Parallel agents are not free: each fork consumes tokens for the inherited context. Apply forking to tasks that are genuinely independent and worth the cost
Fork-by-default turns Claude Code into a practical multi-agent coordinator. Splitting work across parallel agents, collecting clean results, and keeping the orchestrator context readable is now the baseline experience, not an advanced configuration.
Sources: