Anthropic's Claude Code v2.1.140 patches 'silent failures' in agent workflows

Claude Code is not just a chatbot in a terminal anymore. It is an agentic coding tool that reads a repo, runs commands, edits files, and keeps working toward a goal across turns. That is useful — but it also creates a nasty failure mode: the system can look busy, or even look done, while something underneath quietly broke. Anthropic’s May 12 release, Claude Code v2.1.140, is basically a cleanup pass on exactly that class of problem. ### What changed in this release? The official changelog for v2.1.140 is short, but the pattern is obvious once you read it end to end. Anthropic fixed `/goal` “silently hanging” under certain hook settings, fixed `claude --bg` failing with “connection dropped mid-request” as the background service idled out, fixed startup failures on locked-down enterprise machines, and added warnings when plugin folders are being silently ignored because of `plugin.json` settings. (code.claude.com) ### Why call these “silent failures”? Because the dangerous version of failure is not a red error banner. It is a workflow that appears to keep going while the real work has stalled, been skipped, or never started. A hanging `/goal` indicator is one example — the user sees an unresolved state instead of a clear stop. A dropped background connection is another — the agent disappears mid-task. A silently ignored plugin folder is maybe the purest case, because the configuration looks valid until you notice the expected behavior never happened. (code.claude.com) ### Why is `/goal` such a big deal? `/goal` was introduced one version earlier, in v2.1.139. It lets Claude keep working until a completion condition is met, including across turns and in remote-control style workflows. That is a big step toward more autonomous agent behavior. But the catch is obvious — once you tell software to keep going until some condition is satisfied, the worst bug is not “task failed.” The worst bug is “task looks in progress forever” or “task says done without proving it.” (code.claude.com) ### What about the routing fix? Anthropic also made Agent tool `subagent_type` matching more forgiving, so values become case- and separator-insensitive — like “Code Reviewer” resolving to `code-reviewer`. That sounds small, but it is really a reliability patch. Agent systems break in boring ways — a mislabeled role, a config mismatch, a naming convention drift. Relaxing that parser reduces the odds that a subagent fails to launch because a human used the wrong capitalization. (code.claude.com) ### Why do background sessions matter so much? Because background execution is where trust gets expensive. If an agent is running out of band — maybe editing code, polling tasks, or waiting on tools — the user cannot constantly supervise it. So a dropped connection or startup failure is not just a bug. It breaks the whole promise of delegation. Anthropic’s changelog shows two separate fixes here: one for mid-request drops during idle exit, and one for slow startup on enterprise-secured machines. That tells you the company is hardening the boring infrastructure that agent workflows depend on. (code.claude.com) ### Is this bigger than one patch? Yes. Anthropic’s own product page says Claude Code now reads codebases, runs tests, uses tools like GitHub CLI, and is already heavily used inside Anthropic itself. Once a product moves from “answer my question” to “go do the work,” reliability bugs stop being cosmetic. They become execution bugs — wasted compute, half-finished changes, misleading status, and false confidence. ### So what is the real story? (code.claude.com) The real story is not that Claude Code got prettier or slightly faster. It is that Anthropic is patching the trust layer. Agent products live or die on whether users believe “done” means done. v2.1.140 is a small release, but it points at the hard part of the whole category — making autonomous software fail loudly, specifically, and in time for someone to do something about it. (anthropic.com)