CLAUDE CODE 2.1.92 SHIPS FAIL-CLOSED POLICY, AWS BEDROCK SETUP WIZARD, AND CLEARER COST TELEMETRY; ANTHROPIC DETAILS A THREE-AGENT HARNESS FOR LONG-RUNNING WORK
Anthropic updated Claude Code with stronger governance, easier AWS Bedrock setup, and better cost visibility, while sharing a concrete pattern for long-running ...
Anthropic updated Claude Code with stronger governance, easier AWS Bedrock setup, and better cost visibility, while sharing a concrete pattern for long-running multi-agent sessions.
The 2.1.92 release adds a fail-closed policy switch for managed settings, an interactive AWS Bedrock setup wizard, per-model and cache-hit breakdowns in /cost, prompt-cache expiry hints, faster large-file writes, and a pile of CLI/session fixes according to the release notes Claude Code 2.1.92.
Separately, Anthropic outlined a three-agent harness that splits planning, generation, and evaluation, uses context resets with structured handoffs, and employs a dedicated evaluator (with Playwright MCP) to grade outputs during multi-hour runs InfoQ coverage.
If you’re formalizing agent workflows, the Agentic Engineering skill echoes these ideas with evaluation-first cycles, granular task decomposition, and cost-aware routing across Haiku/Sonnet/Opus Agentic Engineering.
Fail-closed policy fetching, cost telemetry, and a guided Bedrock setup reduce risk and surprises in regulated or budget-sensitive environments.
A concrete multi-agent pattern helps keep long-running coding or data tasks coherent and auditable.
-
terminal
Enable forceRemoteSettingsRefresh, then simulate a policy endpoint outage; verify the CLI refuses to start and surfaces clear errors.
-
terminal
Run the same task across Haiku/Sonnet/Opus with and without prompt cache; compare /cost per-model and cache-hit breakdowns to set budgets.
Legacy codebase integration strategies...
- 01.
If you automate via tmux or Homebrew, validate that the subagent spawn fix and cask channel names (claude-code vs. claude-code@latest) don’t break scripts.
- 02.
Use the Bedrock wizard to standardize region/model pinning, and align IAM and credential rotation with existing AWS controls.
Fresh architecture paradigms...
- 01.
Adopt the three-agent harness pattern with explicit evaluator criteria to gate merges for long-running generation or refactors.
- 02.
Start on Bedrock, pin models, and instrument /cost from day one to prevent silent cost creep.