Claude Code raises its agent fan-out cap to 256

Claude Code 2.1.269 raises Workflow's agent concurrency cap to 256 through a new environment variable, per the CHANGELOG. CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS accepts any value from 1 to 256, for inference-bound fan-outs, jobs where each of dozens of agents needs its own model call. That lets more of them run in one pass instead of queuing across batches.

The same release carries two prompt-cache fixes. One stops a cache from invalidating in full after output hits its token limit, so only the affected part drops. The other changes how context gets re-sent when a session interrupted mid-thought is resumed.

A higher cap doesn't make a fan-out faster on its own. The cap was never what stopped a fan-out from being useful. Merging every agent's output into one coherent result is what slows a large fan-out. Push concurrency to 256 without changing how results get combined, and that merge step is where the job stalls.