Breaking

Anthropic Overhauls Claude Code's /fork Command to Spawn Parallel Background Agents

July 21, 2026 at 09:15 EDT

AI Agents
Software Dev & Coding

Anthropic has reworked the /fork command in Claude Code, allowing developers to spin off a fully context-aware agent that runs a side task in the background while the main session keeps moving. The change, shipped in version 2.1.212 around July 17, 2026, is designed to enable true parallel agent work without interrupting a developer's flow.

July 17, 2026 · Claude Code v2.1.212

Claude Code reworks /fork for true parallel agents

The command now spins off a fully context-aware agent into a background session — letting the main thread keep moving while a side task runs in isolation and returns a summary.

~90%

Token savings reported via prompt-cache sharing for child agents

200

Subagent spawns per session cap (also WebSearch calls)

2 min

MCP tool calls beyond this are auto-moved to background

Token cost of a child agent

With cache sharing (CLAUDE_CODE_FORK_SUBAGENT=1), a forked agent runs at a fraction of the tokens

Without sharing

100%

With cache sharing

~10%

Up to 90% fewer tokens for child agents — pairing genuine parallelism with sharply lower cost.

How the reworked /fork flows

/fork
Copies full conversation context

→

Background session
Runs side task in isolated window

→

claude agents
Appears as its own line, monitorable

→

Summary returns
Delivered back to main thread

The old in-session subagent behavior is now split out into a separate /subtask command.

Operational guardrails

Keeping runaway agents in check while enabling parallel work

Session caps

WebSearch and subagent spawns limited to 200 each — configurable via CLAUDE_CODE_MAX_* and resettable with /clear.

Auto-backgrounding

Long MCP tool calls (>2 min) move to the background automatically, adjustable via CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS.

Why it matters

Full context inheritance makes it easier to evaluate decisions from multiple angles, while the mix of token savings and parallelism suits long workflows and parallel research — a shift toward multi-agent coding that is both safer and more manageable.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.