CLAUDE CODE V2.1.97 TIGHTENS SAFETY, FIXES RELIABILITY PAIN POINTS, AND SURFACES LIVE SUBAGENTS
Anthropic shipped Claude Code v2.1.97 with stronger permission hardening, better retry logic, MCP leak fixes, and an indicator for live subagents. The official...
Anthropic shipped Claude Code v2.1.97 with stronger permission hardening, better retry logic, MCP leak fixes, and an indicator for live subagents.
The official release adds a live "● N running" marker in /agents for agent types with active subagents, a hardened Bash tool, and a fix where --dangerously-skip-permissions could silently downgrade to accept-edits mode. It also restores sanity to 429 handling with minimum exponential backoff and fixes OAuth metadata refresh and MCP HTTP/SSE buffer leaks that could grow ~50 MB/hr (v2.1.97 notes, all releases).
This follows community reports of regressions that hurt multi-file workflows in Feb–Mar deep-dive and frustration on YouTube complaint, even as earlier updates claimed speed gains v2.1.94 overview. If you run coding agents, revisit sandbox boundaries; security folks highlight blast-radius control as the main defense for acting agents sandboxing guide.
Third-party coverage also says Anthropic is rolling out managed agents that create and coordinate subagents, and the CLI now visibly tracks running subagents news.
Safer defaults and backoff fixes reduce the risk of runaway edits, noisy retries, and credential issues in agent-driven dev workflows.
Live subagent visibility plus media coverage of managed agents signal multi-agent orchestration moving into day-to-day developer tools.
-
terminal
Upgrade to v2.1.97 in a throwaway repo; probe permission prompts (env var access, redirects), verify sandbox boundaries, and observe the /agents live subagent counter.
-
terminal
Simulate rate limits and OAuth refresh flows; confirm exponential backoff on 429 and correct use of oauth.authServerMetadataUrl after restart; monitor MCP SSE reconnect memory.
Legacy codebase integration strategies...
- 01.
Pin/rollback CLI versions per repo; validate managed-settings and permissions.additionalDirectories behaviors across sessions to avoid surprise access changes.
- 02.
Harden agent sandboxes and shells; update IdP metadata URLs for ADFS-like setups and watch agent spawn counts to prevent resource contention.
Fresh architecture paradigms...
- 01.
Design for least-privilege agent sandboxes from day one; isolate build, test, and deployment tools behind scoped capabilities.
- 02.
Instrument agent/subagent metrics (spawn counts, rate limit events, tool calls) and alert on anomalies before production use.