Claude Code 2.1.302: fork subagents by default and MCP 2026-07-28 tool fixes

Introduction

Mid-October Claude Code changes are less about a new model ID and more about default multi-agent topology and MCP reliability. Defaults that previously differed by client now converge, which can change concurrency, spend, and failure modes overnight if fleets upgrade silently.

Primary source: Claude Code changelog entry for 2.1.302. Adjacent: Claude Managed Agents dynamic workflows, MCP EMA, agent observability stack, and Reflection Beam open agents.

Fork subagents become the default

Fork subagents (`subagent_type: "fork"`) were already default in the terminal. 2.1.302 extends that default to `claude -p`, the Agent SDK, desktop, and VS Code. That means headless and IDE sessions that previously stayed single-threaded may spawn forked workers unless `CLAUDE_CODE_FORK_SUBAGENT=0` is set. `claude mcp serve` remains an exception and never offers fork subagents.

MCP 2026-07-28 negotiation fix

Servers that negotiate protocol version 2026-07-28 but omit `resultType`, `ttlMs`, or `cacheScope` previously showed no tools or failed every tool call. 2.1.302 fixes that negotiation path. Enterprises that saw “empty MCP” after adopting 2026-07-28 servers should upgrade clients before blaming the server implementation.

• Retest every approved MCP server after upgrading Claude Code and the Agent SDK.
• Confirm VS Code MCP sign-in still works for servers whose URLs include query strings (also fixed in this release).
• Keep EMA / admin allowlists in sync so restored tools do not suddenly widen production access.

Hooks, auto mode, and OpenTelemetry

Prompt and agent hooks now queue so at most 12 handlers of one event run at a time. That reduces stampedes when many plugins register the same hook. `CLAUDE_CODE_AUTO_MODE_MODEL` selects the model for the client-side auto-mode classifier without changing server-side auto mode. OpenTelemetry `tool_decision` events regain `classifier_category` in more auto-mode blocks and keep `tool_use_id`, `tool_source`, and `tool_parameters` when attribute count limits are low.

What this means for developers

• Pin Claude Code / Agent SDK to 2.1.302+ in CI images and document whether fork is intentional.
• If a workflow must stay single-agent, set `CLAUDE_CODE_FORK_SUBAGENT=0` explicitly rather than relying on old client defaults.
• Re-run MCP smoke tests for every server on the 2026-07-28 protocol.
• Verify OTel dashboards still show tool decisions after the attribute fixes.
• Expect desktop/VS Code subagents to stay backgrounded; adjust UX and monitoring accordingly.

What this means for businesses

• Default forking is a governance event: update runbooks, spend alerts, and sandbox requirements the same week as the client bump.
• MCP empty-tool incidents may have been client bugs; schedule a coordinated client upgrade rather than rewriting every server.
• Hook concurrency caps reduce surprise load from plugin packs, but teams should still inventory which hooks fire on agent start.
• Pair this release with Managed Agents workflow limits so interactive Claude Code and background Managed Agents share one concurrency story.

One-week fleet checklist

1. 01
   Inventory clients

   List Claude Code CLI, desktop, VS Code, and Agent SDK versions in CI and developer machines.
2. 02
   Decide fork policy

   Enable default forks for pilots that need parallelism; set CLAUDE_CODE_FORK_SUBAGENT=0 where single-thread is required.
3. 03
   Upgrade and retest MCP

   Bump to 2.1.302+, then smoke every allowlisted MCP server for tool discovery and one real write path.
4. 04
   Reset spend alerts

   Compare token burn on three gold tasks before vs after fork defaults.
5. 05
   Confirm OTel

   Validate tool_decision attributes land in the existing collector with classifier_category present when expected.

Conclusion

Claude Code 2.1.302 is an agent-operations release: fork defaults converge across clients, nested agent foregrounding is explicit, and MCP 2026-07-28 negotiation stops looking like a broken server. The useful move is a controlled upgrade with an explicit fork policy and a full MCP retest, not a silent auto-update across the fleet.

Sources: Claude blog