GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 827 -5.3%
QWEN-38X 824 —
CL-OP55X 820 —
GPT-6A 820 —
GROK-46H 820 -5.2%
GLM-5 784 -8.4%
KIMI-K3X 742 -8.4%
CL-FAB5H 742 -5.7%
CL-OP5H 718 -6%
CL-OP5X 708 -18.2%
CL-OP46H 696 -6.2%
CL-OP47H 688 -6.1%
GEM-38FH 677 +0.1%
GEM-37FH 655 -24.3%
GPT-56S 619 —
GPT-55H 580 —
CL-OP47 579 -0.7%
INKL 531 —
GEM-31P 512 —
GEM-3P 498 —
CL-OP46 496 —
CL-OP48 489 -0.2%
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 827 -5.3%
QWEN-38X 824 —
CL-OP55X 820 —
GPT-6A 820 —
GROK-46H 820 -5.2%
GLM-5 784 -8.4%
KIMI-K3X 742 -8.4%
CL-FAB5H 742 -5.7%
CL-OP5H 718 -6%
CL-OP5X 708 -18.2%
CL-OP46H 696 -6.2%
CL-OP47H 688 -6.1%
GEM-38FH 677 +0.1%
GEM-37FH 655 -24.3%
GPT-56S 619 —
GPT-55H 580 —
CL-OP47 579 -0.7%
INKL 531 —
GEM-31P 512 —
GEM-3P 498 —
CL-OP46 496 —
CL-OP48 489 -0.2%
← Back to feed

Anthropic's 'Dreaming' Lets Claude Agents Learn From Their Own Mistakes Between Sessions

Anthropic used its second annual Code with Claude developer conference in San Francisco this week to announce a new capability for Claude Managed Agents: a system called dreaming that lets deployed agents learn from mistakes across sessions without requiring manual retraining or human intervention.

The mechanism works by having agents replay failed or suboptimal task sequences during low-utilisation periods, adjusting internal heuristics based on what went wrong. Anthropic described it as analogous to sleep-based memory consolidation — agents process experience asynchronously rather than only at runtime.

The Managed Agents Platform Gets Richer

Claude Managed Agents launched earlier this year at $0.08 per compute-hour. The dreaming update follows the persistent memory feature that shipped in public beta last month. Rakuten, an early production user, reported a 97% reduction in task errors after enabling persistent memory. Anthropic didn’t publish comparable numbers for dreaming at launch, saying early deployments across finance and legal verticals are still being evaluated.

The conference also included a compute announcement: Anthropic confirmed a deal with SpaceX to access the full capacity of the Colossus 1 data centre in Memphis, Tennessee — a 220,000-GPU facility. The deal doubled Claude Code usage limits for all tiers.

Boris Cherny: AI Writes All the Code Now

Boris Cherny, head of Claude Code, described a shift in how he works that has become common inside Anthropic’s own engineering teams.

“About six months ago what happened is all the code that I used to have written by hand now Claude writes. And I just prompt Claude. So I talk to Claude and I’m like, hey, let’s build this feature. It builds a feature and it tests it. And then it shows me. And I’m like, yeah, that’s good. Or no, wait, make this change.”

Cherny also told Business Insider he wants a new term to replace “vibe coding” — arguing the phrase undersells the degree of engineering judgment still required in effective human-AI coding loops.

What Dreaming Changes for Agentic Deployments

The practical implication is that long-running agent deployments now improve over their lifetime without requiring explicit retraining jobs or periodic human review. For enterprise use cases — financial reconciliation, compliance screening, document processing — this closes a gap that required ongoing human supervision to manage error accumulation.

The dreaming capability is available in Claude Managed Agents starting from the Anthropic developer console. It requires persistent memory to be enabled; agents need a record of past sessions to replay against.

Pricing for dreaming is included in the existing $0.08/hour compute rate during active tasks. Dreaming sessions during idle periods are billed at $0.02/hour.