Anthropic adds "dreaming" to Claude Managed Agents and doubles Pro/Max five‑hour limits

Anthropic unveiled “dreaming” at its Code with Claude developers’ conference in San Francisco, rolling out a research‑preview capability for Claude Managed Agents that periodically reviews recent sessions and memory stores to surface and preserve high‑signal patterns across agents. The company framed the feature as a cross‑agent memory curation process rather than a chat‑level compaction step, saying it is intended to keep important context available for long‑running tasks and multi‑agent workflows — an issue that otherwise causes key information to fall out of limited model context windows.

Developers who coordinate agents over minutes or hours are the immediate target for the feature. Dreaming runs as a scheduled review that analyzes past sessions and memory stores, identifies specific items worth retaining, and restructures memory so the retained items remain high‑signal over time. Users can choose an automatic process or inspect proposed memory changes manually. Anthropic emphasized that dreaming is limited to Managed Agents on the Claude Platform. a pre‑built, configurable agent harness running in managed infrastructure, designed for scenarios where multiple agents collaborate toward endpoints over extended interactions.

Anthropic gave examples of the kinds of cross‑agent patterns dreaming is meant to surface — recurring mistakes, workflows agents tend to converge on, and team‑shared preferences — and said the curation and restructuring of memory should improve long‑running work and multi‑agent orchestration where context retention and pattern recognition matter. The company contrasted this cross‑agent approach with single‑agent memory handling, arguing that patterns spanning multiple agents won’t reliably appear to any one agent working in isolation.

Alongside dreaming, Anthropic said two previously disclosed research previews — outcomes and multi‑agent orchestration — are being made more broadly available. It also announced it will double the five‑hour usage limits for subscribers on its Pro and Max plans, a response to user frustration as its compute infrastructure has struggled to keep pace with demand. Anthropic did not provide a timetable for a general rollout, and positioned these changes and expanded preview access as interim measures to ease developer pain points while it grows compute capacity.