Anthropic reveals 'dreaming' feature and multi-agent orchestration at Code w/ Claude preview

Anthropic’s latest Claude push is not really about chat. It’s about runtime. The company used its Code with Claude event on May 7 to show new features for Claude Managed Agents — including “dreaming,” outcomes, webhooks, and multi-agent orchestration. The big idea is simple: stop treating the model as a smart textbox and start treating it like software that can take work, wait, remember, coordinate, and report back. That matters because long-running agent workflows usually break in boring ways. They lose context. They need too much babysitting. They can do one clever step, but not ten dependable ones. Anthropic is trying to close that gap. ### What is “dreaming,” exactly? Dreaming is Anthropic’s name for a memory pass that happens after work, not during it. The agent reviews recent sessions, looks for patterns worth keeping, and stores useful lessons for future tasks. Ars Technica’s description is basically “reflection plus memory” — not sleep, not consciousness, just a structured way to decide what should stick. ### Why does that matter for coding agents? Because coding work is repetitive in the worst way. The same repo conventions, the same failed tests, the same deployment gotchas keep coming back. A normal chat session forgets all of that unless you keep pasting context back in. Dreaming is Anthropic’s attempt to let agents build a working memory over time so they make fewer repeat mistakes and need less manual steering. ### What are outcomes and webhooks doing here? They make the agent act more like infrastructure. Outcomes give developers a clearer way to specify what “done” means for a task. Webhooks let outside systems kick off or react to agent work — think repo events, CI signals, or app-side triggers. That shifts Claude from “ask me something in a terminal” to “plug me into the flow of engineering events.” ### And multi-agent orchestration? That is the other big piece. Anthropic has been moving this way for a while. Its own research system already uses an orchestrator-worker setup, where one lead agent delegates work to specialized subagents in parallel. Now that pattern is being exposed more directly to developers building with Managed Agents. One agent plans, others execute, and the system coordinates the handoffs. ### Why package this under Managed Agents? Because Anthropic has been arguing that the harness matters as much as the model. In April, it described Managed Agents as a hosted service for long-horizon work with interfaces that stay stable even as the underlying harness changes. That is a very specific bet: developers do not want to rebuild orchestration logic every time model behavior improves. They want a control plane that survives model churn. ### Is this separate from Claude Code? Not really — it looks more like the next layer above it. Claude Code started as a terminal-based research preview for delegating substantial engineering tasks. Since then Anthropic has added subagents, hooks, plugins, sandboxing, and more autonomy features. This week’s preview extends that trajectory from “agentic coding tool” toward “agentic engineering system.” ### So what changed today? The change is product shape. Anthropic is no longer just selling a model that happens to code well. It is assembling the scheduling, memory, triggering, and coordination pieces needed for async software work. That is a different category — closer to an engineering workflow layer than a chatbot with tools. AI coding is not just who has the smartest model. It is who builds the most reliable operating layer around that model — the part that remembers, waits, delegates, and plugs into real systems. If that layer works, “use Claude” starts to mean “wire Claude into the team’s workflow,” not just “open a chat window.”