Anthropic lets Claude agents 'dream' to get smarter

What if your AI agents actually got smarter while you slept?

Anthropic just rolled out three major features for Claude Managed Agents: dreaming (in research preview), along with outcomes, multiagent orchestration, and webhooks (all in public beta). The update tackles the three biggest headaches in running production AI agents: amnesia between tasks, sloppy outputs, and bottlenecks from forcing one agent to do everything alone.

Dreaming away bad habits

Dreaming acts as a scheduled maintenance routine for agent memory. Instead of merely recording notes on the fly, dreaming analyzes past sessions and stored context between jobs to extract high-level patterns. It flags recurring errors, identifies common team preferences, and weeds out noise from memory stores.

Developers can let dreaming update agent memory automatically or require human approval before changes stick. Legal AI startup Harvey has been testing the feature for long-form drafting and document generation, reporting that agents successfully retained file workarounds and tool patterns, boosting completion rates by roughly 6x.

Grading rubrics and team captains

For quality control, outcomes lets developers define a strict rubric of what "good" looks like. A dedicated grader evaluates the agent's work in an entirely separate context window, keeping it uninfluenced by whatever reasoning the worker agent used. If the output misses the mark, the grader tells the agent what to fix for another pass. Anthropic reports outcomes improved task success by up to 10 points on internal benchmarks, bumping docx generation success by 8.4% and pptx by 10.1%. Wisedocs used it to speed up document reviews by 50%.

Meanwhile, multiagent orchestration lets a lead agent split massive tasks across multiple specialized subagents working in parallel on a shared filesystem. You can mix and match models: Spiral by Every runs its coordinator on Haiku to field prompts, delegates parallel writing drafts to multiple Opus instances, and uses outcomes to verify the prose against editorial guidelines. Netflix uses the same setup to churn through logs from hundreds of builds at once.

Everything logs to the Claude Console, tracking exactly which agent ran what tool and why.

Why it matters

We think Anthropic is smartly addressing the biggest barrier to enterprise agent adoption: unpredictable execution.

Until now, teams had to stitch together fragile LangChain loops or roll their own evaluation scripts to keep agents from delivering hallucinated junk. Anthropic is moving those patterns into native cloud infrastructure. Isolating the grader into its own context window is a particularly sharp move, dodging the common pitfall where an agent simply convinces itself that its own mistake is correct.

That said, autonomy isn't free. Running parallel subagents paired with multi-turn rubric grading will burn through tokens fast and add noticeable latency. While routing simple orchestration through lightweight models like Haiku helps keep costs sane, developers will need to carefully weigh the token bill against the hours saved on manual review.