Claude Code v2.1.232｜サブエージェントの「fork」がデフォルト有効にClaude Code v2.1.232 enables fork-type subagents by default, allowing them to…

Anthropic has shipped Claude Code v2.1.232, and the release notes highlight a small but meaningful change to how the tool handles delegated work: subagents of the "fork" type are now enabled by default. For developers who lean on Claude Code's agentic capabilities to break large tasks into parallel pieces, the shift affects how quickly and cheaply those parallel branches spin up.

The change centers on how a subagent is initialized. In Claude Code, the main agent can spawn subagents to handle discrete pieces of work, and each subagent is described by a subagent_type. Previously, launching that kind of helper generally meant starting with a comparatively fresh context. With v2.1.232, a subagent declared with subagent_type set to "fork" inherits the parent's entire conversation history along with its prompt cache. Because this behavior is now the default rather than an opt-in setting, workflows that fan out into multiple agents should pick it up automatically.

Inheriting the prompt cache is the technically interesting part. Anthropic's models support prompt caching, a mechanism that stores the processed representation of a large, stable block of input so it does not have to be recomputed on every call. When a forked subagent reuses the parent's cache, it avoids re-reading and re-tokenizing the shared context from scratch, which is where much of the latency and token cost in long agent sessions tends to accumulate. Carrying over the full conversation context also means the subagent already "knows" what the parent was working on, so it can begin its task without the parent having to restate the relevant background. The stated result, according to the changelog summary, is improved efficiency and faster response times for multi-agent, parallel workflows.

It helps to place fork alongside the alternative. A subagent that starts without inherited context is effectively a clean slate: it can be given a narrow, self-contained brief and will not be influenced by unrelated details in the parent conversation, which can be useful for isolation and for keeping token usage predictable. A forked subagent trades that isolation for continuity and speed, since it begins with everything the parent already established. Making fork the default suggests Anthropic sees the continuity-and-speed tradeoff as the more common need for typical Claude Code usage, though users who want stricter separation between tasks presumably retain the ability to choose other subagent behaviors.

For readers less familiar with the surrounding tooling, Claude Code is Anthropic's command-line coding agent, positioned to read a codebase, plan changes, run commands, and iterate on tasks with relatively little step-by-step supervision. Subagents fit into that picture as a way to decompose a big job—say, refactoring several modules or investigating multiple files at once—into concurrent units of work that report back to a coordinating agent. Faster and cheaper subagent startup compounds across those parallel branches, so even a modest per-branch improvement can matter when many are running.

The move also reflects a broader industry pattern. Competing agentic coding tools such as GitHub Copilot's agent features, Cursor, OpenAI's Codex-style tooling, and open frameworks like OpenHands have all been experimenting with multi-agent orchestration and context reuse as ways to control cost and latency. Prompt caching in particular has become a common lever across model providers for reducing the expense of repeatedly sending large, unchanging context, and applying it to subagent creation is a logical extension of that idea within an agent framework.

A few caveats are worth keeping in mind. This item comes from a blog that translates and annotates Anthropic's official Claude Code changelog, so the framing here follows that source rather than independent testing. The practical gains from forked subagents will likely depend on how large the shared context is and how many subagents a given workflow spawns; sessions with small contexts may see less benefit. Teams that relied on subagents starting fresh should also verify that the new default matches their expectations, since inheriting the full parent conversation changes what information a subagent can access. As always with point releases, reviewing the changelog before upgrading remains the safest way to confirm the behavior.