Anthropic shipped something called Dreaming for Managed Agents this week. It’s a scheduled background process that runs between sessions — the agent reviews its own past conversation transcripts, extracts patterns, and writes learnings into memory. No human in the loop unless you want one.

The framing that stuck with me: individual sessions are blind to cross-session patterns. A support agent won’t notice it made the same classification error 12 times this month. Dreaming is designed to surface exactly that kind of signal.

It ships alongside Outcomes (automated output grading against developer-defined rubrics) and multi-agent orchestration (coordinator + up to 20 parallel subagents, now in public beta). The three are meant to work as a loop: orchestration decomposes work, Outcomes grades it, Dreaming remembers the failures.

Still in research preview, not GA.

A few things I’m genuinely uncertain about:

The “automatic” mode lets the agent write directly to its own memory without approval. That’s a meaningful amount of autonomy over its own behavior. How do you audit what it’s actually learning? If it develops a subtly wrong heuristic over three months of self-reinforcement, how do you catch that before it’s deeply embedded?

Also curious about the human-review mode in practice — if you’re approving every proposed memory update, does that scale? Or does it become a bottleneck that defeats the purpose?

For those building on Managed Agents or similar systems: are you thinking about self-improvement loops as a feature you want, or a risk you’d rather control tightly? And does the “agent with three months of experience vs. freshly deployed agent” framing change how you think about agent versioning and rollbacks?