Anthropic pushes into workflows
Anthropic is trying to move from selling models to owning enterprise workflows with a new Managed Agents push aimed at production-ready tooling. At the same time it previewed a powerful cybersecurity model — Mythos — to about 40 security firms to patch vulnerabilities, but has withheld public release over misuse concerns. (nationaltoday.com, )
Anthropic spent this week pushing in two directions at once: it launched a service called Claude Managed Agents for companies that want artificial intelligence workers in production, and it separately showed a new model called Claude Mythos Preview to a small security group instead of releasing it widely. (anthropic.com) (red.anthropic.com) The first move is about who owns the job after a model answers a prompt. A chatbot gives you words, but an agent is supposed to keep going for minutes or hours, call tools, recover from errors, and finish a task like a junior employee following a checklist. (anthropic.com) (wired.com) Anthropic says Managed Agents is a hosted service for “long-horizon” work, which is industry shorthand for tasks too long and messy for a single back-and-forth chat. The company says customers can use it to build and deploy agents at scale without wiring up all the infrastructure themselves. (anthropic.com) (siliconangle.com) That sounds technical, but the business point is simple. Anthropic does not just want to sell raw model access by the token if it can also sell the operating layer that decides what tools the model can use, how long it runs, and how it hands work back to a human. (wired.com) (anthropic.com) Anthropic’s own engineers described the problem as “decoupling the brain from the hands.” In plain English, the model is the brain, while the harness is the scaffolding around it that gives it tools, memory, permissions, and rules for taking actions on real systems. (anthropic.com) Anthropic argues those harnesses go stale fast as models improve, because engineers hard-code assumptions about what the model cannot do and then have to keep rewriting them. Managed Agents is supposed to hide more of that churn so customers can plug in a stable interface while Anthropic updates the moving parts underneath. (anthropic.com) The second move came a day earlier and pointed at a much riskier market. Anthropic said Claude Mythos Preview is its most capable frontier model so far, and its system card says the model shows a “striking leap” over Claude Opus 4.6 on many evaluations. (anthropic.com) Anthropic did not put Mythos on a normal public menu because the company says the model is unusually strong at finding and exploiting software vulnerabilities. Instead it created Project Glasswing and limited access to roughly 40 partners working on defensive cybersecurity. (anthropic.com) (red.anthropic.com) (theguardian.com) Project Glasswing is basically a controlled test: let trusted security firms use the model to patch holes before attackers find them. Anthropic says the partner list includes companies such as Apple, Amazon Web Services, Cloudflare, CrowdStrike, Datadog, Palo Alto Networks, and Microsoft. (anthropic.com) (cnbc.com) Put together, the two announcements show Anthropic trying to climb the stack in opposite directions. Managed Agents moves upward from models into enterprise workflow software, while Mythos moves downward into one of the most sensitive technical jobs a model can do: reading code, spotting flaws, and suggesting fixes before those flaws turn into break-ins. (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) That leaves Anthropic in a position a lot of artificial intelligence companies want but few have yet secured. If customers let Anthropic run the agent layer for everyday work and trust its restricted models for high-stakes security work, the company is no longer just selling a smarter autocomplete box; it is renting out part of the company operating system. (wired.com) (siliconangle.com)
