Anthropic pilots 'Dreaming' and adds Agent View to Claude Code to boost agent workflows

Anthropic is trying to make Claude feel less like a chatbot and more like a working system. Not one model answering one prompt, but a bundle of agents that can plan, branch, execute, and keep going for a while. The news around “Dreaming” and Agent View fits that bigger shift. So does AWS making Claude Platform on AWS generally available on May 11. But the twist is that Anthropic is shipping all this while also warning that its newest models are getting dangerous enough to matter for real cybersecurity. ### What is Anthropic actually building? Basically, a more autonomous Claude stack. Claude Code already goes beyond autocomplete — it reads a codebase, edits files, runs tests, and ships changes. Anthropic has also been building hosted Managed Agents for long-horizon work, plus multi-agent systems where one lead agent spins up parallel subagents to explore different paths at once. Agent View makes sense in that context. If you have several agents working in parallel, you need a way to see what each one is doing without getting lost in a terminal scrollback. (aws.amazon.com) ### Why does multi-agent work matter? Because a lot of useful work is bottlenecked by breadth, not just raw intelligence. Anthropic’s own engineering write-up on Research spells this out pretty clearly: one agent can get path-dependent and miss whole branches, while parallel subagents can explore independent directions and then compress the useful bits back to a lead agent. That is the same basic logic behind agent workflows in coding — one branch investigates a bug, another writes tests, another checks docs, and a coordinator pulls the pieces together. “Dreaming,” whatever Anthropic’s exact internal label means in this pilot, appears to sit in that family of tricks for improving task completion by giving the system more room to think, branch, or recover. (anthropic.com) The point is not chat quality. The point is finishing harder jobs. ### Why add AWS now? Because enterprises buy plumbing, not just models. Claude Platform on AWS gives teams the native Claude Platform through an existing AWS account, with IAM, CloudTrail, and consolidated billing already in place. AWS says it is the first cloud provider to offer Anthropic’s native platform experience directly, including APIs, console access, Managed Agents in beta, web search, web fetch, code execution, files API, Skills, and MCP connector support. That lowers one of the biggest adoption frictions — procurement and governance overhead. (anthropic.com) ### What’s the catch for security teams? The catch is that Anthropic is not pretending these systems are harmless. In March, it said Claude Code users approve 93% of permission prompts, which is why it built auto mode as a middle ground between constant manual approvals and the very risky “skip permissions” approach. Anthropic’s examples of agent mistakes were not theoretical — deleting remote git branches, exposing an auth token, and attempting production database migrations. More autonomy means more chances to do the wrong thing faster. (aws.amazon.com) ### Why is Mythos the real backdrop here? Because Mythos changes the tone from “useful automation” to “serious capability.” Anthropic says Mythos Preview could identify and exploit zero-day vulnerabilities in every major operating system and every major web browser during testing, and that over 99% of the vulnerabilities it found are still undisclosed. It launched Project Glasswing to push those capabilities toward defensive security work with critical software organizations. So the company is advancing agent workflows at the same time it is telling customers that frontier models are crossing into offensive-cyber territory. (anthropic.com) ### So what should enterprises take from this? Anthropic’s pitch is getting clearer. Claude is becoming a platform for delegated work — coding, research, document-heavy operations, and domain-specific agent flows — not just a model endpoint. But the company’s own safety work says the same autonomy that makes these products valuable also makes them risky. If you are adopting Claude for agent workflows, the real product is not just the model. It is the control layer around the model. (red.anthropic.com) ### Bottom line? Anthropic is productizing the age of agents. But it is also quietly admitting the age of agent security is already here. (red.anthropic.com) (aws.amazon.com)