Anthropic’s Forked Subagents Fix Claude Code’s Biggest Context Problem
Wait, more context helps? Anthropic’s Claude Code update says yes. Forked subagents tackle its biggest weakness: helpers losing the plot.
Wait, more context helps? Anthropic’s Claude Code update says yes. Forked subagents tackle its biggest weakness: helpers losing the plot. Instead of stuffing one thread, noisy tool work splits off, then clean answers snap back.
Here is the change. Standard subagents get a compressed handoff. Forked ones inherit the full conversation and instructions. They also share the same prompt cache, so carrying all that history does not automatically mean a bigger bill.
That matters once a session gets nuanced. One workflow built up about 50,000 tokens, then squeezed each normal helper to roughly 2,000. Forks keep the detail, so parallel variants, verification, and follow-up tool steps come back smarter, not flatter.
Use forks when context is the advantage: /fork, /recap, and /btw. You can keep asking inside the branch without polluting the main thread. The twist is practical: more context is better, unless you want a blank reviewer for independent code critique.
Key facts
- Anthropic introduced forked subagents for Claude Code to address limitations of standard subagents.
- Forked subagents inherit the full prior history of the main conversation and its instructions.
- Standard subagents receive a summarized handoff rather than the full conversation context.
- A cited design workflow built up about 50,000 tokens of nuanced discussion before being compressed to roughly 2,000 tokens for each standard subagent.
- In one live example, forked subagents started with about 180,000 tokens because they inherited the existing conversation state.
Why it matters
Forked subagents preserve nuance from long working sessions, improving parallel exploration, verification, and research quality without cluttering the main thread.
The Signal Brief
One sourced idea worth your attention, in your inbox. No noise.