Anthropic Unveils 'Dreaming' for Self-Improving AI Agents Amid 80x Growth

Anthropic introduced a new technique called "dreaming" for its Claude Managed Agents at the company's Code with Claude developer conference in San Francisco. Despite the anthropomorphic name, this isn't about AI having human‑like dreams. It's a scheduled, between‑session process where Claude reviews its recent activity and identifies patterns worth saving into persistent memory.

"Dreaming surfaces patterns that a single agent can't see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across a team," Anthropic said, according to Ars Technica. "It also restructures memory so it stays high‑signal as it evolves."

The feature is launching as a research preview, requiring developers to apply for access. It's limited to Managed Agents — Anthropic's higher‑level alternative to directly using the Messages API — and isn't yet available to all developers.

Most chat‑based AI systems already use compaction: condensing long conversations so irrelevant details are removed while important context is preserved. But compaction is typically limited to a single conversation thread. Dreaming operates across multiple agents and sessions, pulling insights from past interactions that individual agents would miss on their own.

Business Insider describes the technique as built to "refine a system's memory by running evaluations between sessions. It will review old behavior and seek patterns, then help agents establish better ways of working and cut down on mistakes." Users can either let dreaming run automatically or review and approve proposed memory changes manually.

The practical implication: agents that get more accurate and more productive over time as they accumulate experience, rather than starting fresh every session. For builders running long‑running agent workflows, this could dramatically reduce the need to re‑explain context or fix the same mistakes repeatedly.

CEO Dario Amodei used the developer conference to put a number on Anthropic's explosive growth. The company grew 80x in Q1 on an annualized basis, far beyond the 10x growth Anthropic had planned for. "That is the reason we have had difficulties with compute," Amodei said, as reported by CNBC.

Amodei described the growth as "just crazy" and "too hard to handle," expressing hope for "more normal" expansion going forward. The surge has been driven largely by Claude Code, which Amodei said has taken the software development world "by storm."

"Software engineers are the ones who are fastest to adopt new technology," Amodei said. "It's a foreshadowing of how things are going to work across the economy, and how the economy is going to be transformed by AI." Reuters also reported Anthropic's Q1 revenue grew "80x" on an annualized basis, with a slide at the event declaring: "Coding has changed forever. Finance is next."

Alongside dreaming, Anthropic announced three immediate changes to Claude usage limits. Claude Code's five‑hour rate limits are doubling for Pro, Max, Team, and seat‑based Enterprise plans. Peak‑hour limit reductions are being removed entirely for Pro and Max accounts. And API rate limits for Claude Opus models are rising considerably, according to Anthropic.

The expanded capacity is backed by a newly signed deal with SpaceX's Colossus 1 data center, which gives Anthropic access to more than 300 megawatts of capacity and 220,000+ NVIDIA GPUs within the month. Anthropic also announced a raft of other infrastructure deals: up to 5 gigawatts with Amazon, a 5 GW agreement with Google and Broadcom starting in 2027, and a $30 billion Azure partnership with Microsoft and NVIDIA.

Two additional Managed Agents features moved from research preview to public beta: rubric‑style outcomes that guide agent behavior against defined standards, and multi‑agent delegation that lets a primary agent distribute work across specialized sub‑agents.

Dreaming fits into a larger trajectory Anthropic has been telegraphing. Co‑founder Jack Clark recently predicted a 60% chance that frontier AI models will be able to autonomously train their successors by the end of 2028. While dreaming is a much narrower form of self‑improvement — memory curation, not model training — it's a step toward AI systems that get better at their jobs through experience rather than just through new model releases.

Anthropic is reportedly in fundraising talks at a $900 billion valuation, which would make it worth more than OpenAI, according to CNBC. The company's growth curve — from Claude Code's developer adoption to enterprise finance agents to the SpaceX compute deal — suggests it's betting that being the platform for AI‑native software development is worth more than being the best chatbot.

For builders, dreaming signals where agent infrastructure is heading: persistent, learning systems that compound their value over time. If you're building on Claude's API today, dreaming won't affect you directly yet — it's gated behind research preview on Managed Agents. But the direction is clear. The agents of 2027 won't start from zero every session. They'll have memory, pattern recognition, and the ability to self‑correct based on past work.

The 80x growth number also has a practical implication: Anthropic will continue to struggle with capacity. Even with the SpaceX, Amazon, Google, and Microsoft deals, demand is outpacing supply by a factor of 8. If you're building products that depend on Claude API availability, plan for continued rate limiting and occasional degradation during peak hours. The doubled limits help, but they're a bandage on an 80x wound.