Anthropic is rolling out a new capability for Claude Code that runs while the agent is idle, reviewing past work to sharpen its future performance. The feature, called dreaming, scans up to 100 previous sessions during downtime, hunts for patterns the model missed in real time, flags recurring mistakes and rewrites its long-term memory store accordingly.
"Anthropic is teaching Claude to dream", Jaeden Schafer said on the podcast, framing the launch as one of the more underrated shifts in agent design this year. The pitch is that Claude Code gets better at the work a developer has already given it without having to be retrained, simply by consolidating what it has already seen.
The mechanic mirrors how human memory works overnight. Schafer noted on the show that this is "not training it's just memory consolidation", and that Anthropic is positioning the behavior as the AI analogue of what the brain does during sleep. For coding agents that grind through repeated edits across the same files, the implication is fewer repeated dead ends.
Key facts
- 01Anthropic's new 'dreaming' feature lets Claude Code review up to 100 previous sessions during idle downtime to spot patterns and recurring mistakes.
- 02An accompanying Outcomes Plus feature, which uses a separate grader agent to score task completion, improved task success rate by up to 10 points over Anthropic's standard prompting loop.
- 03Dreaming, Outcomes Plus, and multi-agent orchestration are all now in public beta on the Claude platform.
- 04The process is memory consolidation rather than retraining, with Anthropic comparing it to what the human brain does during sleep.
That repetition is a familiar pain point. Schafer described watching Claude Code chew through tasks in his own workflow, trying an approach, failing, trying another, and sometimes hitting the same wall on a later run. Rather than perfecting every prompt up front, he tends to fire off the request and let the agent grind. Dreaming is aimed squarely at that loop, giving the agent a way to remember which paths already failed.
“Anthropic is teaching Claude to dream”— Jaeden Schafer
Anthropic is shipping a companion feature called Outcomes Plus alongside it. In the company's internal testing, Outcomes Plus, which uses a separate grader agent to score task completion, lifted task success rates by up to 10 points over the standard prompting loop. Combined with dreaming, the goal is an agent that is not only more accurate but also cheaper and faster, because it stops re-treading mistakes it has already made.
Schafer argued the efficiency story is the part developers will feel first. The system should be "number one more efficient number two it's going to use less tokens and number three faster", he said, calling speed the gain everyone can get behind. Token cost is a live constraint for teams running agents in long, multi-step coding sessions, and a memory layer that shaves redundant attempts has a direct line to the bill.
Dreaming, Outcomes Plus and multi-agent orchestration are all now in public beta on the Claude platform, giving Anthropic a three-pronged push into longer-horizon agent work at the same time. It is also a competitive signal. Coding agents have become the most contested surface in applied AI, with OpenAI, Google and a wave of startups all chasing reliability on real engineering tasks.
Schafer's read is that Anthropic is setting the pace. He said rivals "have all their work cut out to try to keep up with what Anthropic is putting out there", citing the cadence of agent-focused features the company has shipped recently. If the 10-point success-rate gain holds outside Anthropic's own benchmarks, the pressure on competing coding assistants to add their own memory consolidation layers will only grow.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
