Anthropic pushes AI into workflows

Anthropic is turning Claude from a chat tool into software that runs jobs on its own, starting with a new automation feature called Routines. (code.claude.com) Anthropic’s documentation says a routine is a saved Claude Code setup — prompt, repositories, and connectors — that can run on a schedule, from an application programming interface call, or from GitHub events on Anthropic’s cloud. The feature is in research preview and is available on Pro, Max, Team, and Enterprise plans. (code.claude.com) Because the jobs run on Anthropic-managed infrastructure, the company says they keep working when a developer’s laptop is closed. Its examples include nightly backlog cleanup, pull request review, deploy checks, alert triage, and documentation updates. (code.claude.com) Anthropic made a separate push into enterprise infrastructure on April 8, when it introduced Claude Managed Agents in public beta. The company’s platform documentation says the service lets customers run long-running agents in managed cloud environments instead of building their own agent loop, tool execution layer, and runtime. (platform.claude.com) Anthropic says those managed agents can read files, run shell commands, browse the web, and execute code inside configured containers with server-side event history. The company describes the core building blocks as an agent, an environment, a session, and events. (platform.claude.com) The company is also packaging reusable behavior into what it calls Skills, which Anthropic documents as an enterprise feature that needs governance, security review, and evaluation before broad deployment. Its own checklist warns reviewers to look for bundled scripts, network calls, hardcoded credentials, and instructions that try to bypass safety rules. (platform.claude.com) Anthropic’s engineering team says the goal is to separate the model from the machinery around it, so the interfaces stay stable even as the underlying harness changes. In a post published last week, the company said it built Managed Agents for “long-horizon” work after earlier harness designs had to change as newer Claude models behaved differently. (anthropic.com) That product expansion is landing alongside user complaints that Claude has become less reliable for complex work. Fortune reported on April 14 that developers and power users have accused Anthropic of reduced performance and weak disclosure, while Anthropic leaders disputed the idea that the company had secretly downgraded the models. (fortune.com) VentureBeat reported on April 14 that Anthropic’s managed-agent pitch could also increase dependence on Anthropic’s stack, because customers get convenience in exchange for giving up some control over orchestration. Anthropic’s own docs frame the tradeoff more narrowly, positioning Managed Agents for asynchronous, long-running jobs and the Messages Application Programming Interface for teams that want finer control. (venturebeat.com) (platform.claude.com) Anthropic is now selling two things at once: a model and the workflow around the model. The next test is whether developers trust Claude enough to let it keep working after they close the laptop. (code.claude.com) (fortune.com)