Anthropic engineer: "The agent doesn't remember anything. So we built a second set of agents whose only job is to dream about the first ones." They wait until you log off, then reopen every session you ran, fact-check the first agents, merge the duplicates, and burn anything that went... stale. One dreaming agent per session, up to 100 at once, built to miss nothing. 95% of the tokens come back cached, so an entire memory rewrite costs almost zero. This is the part of the workflow nobody is talking about yet. Watch the full workshop, then save the setup below.show more

the merging step is the hard part nobody talks about. when agent b reviews agent a's work from hours ago and they disagree, who wins? it's semantic merge conflicts. vcs solved this for lines of code but for agent outputs the resolution strategy is wide open

Been doing this for many months now via magic context

This is really good

Aha what an interesting approach. This kind of “swarms” are going to gib rise to so many ZHCs! @anajuliabit , check this out

Great insight. Agent workflows don't end when the task is done. Memory consolidation and knowledge refinement are becoming core parts of the stack.

The real story here isn't sci-fi vibes, it's context rot. Stuffing massive chat logs into a live loop kills reasoning. Offline consolidation is how we keep our Gemini 3.5 Flash loops fast and cheap.

Agents that dream while you sleep. The workflow just got a second shift

Background agents that audit other agents while you sleep is architecturally interesting but doubles your compute spend and adds a whole new failure mode: what if the auditor hallucinates a false positive?

@heydittoai

@garrytan’s gbrain has had this for months. Welcome to the club.

why wait until you log off? nobody logs off out of claude code

that concept of recursive dreaming agents is actually wild

this is literally just digital sleep cycles for ai

nothing costs nothing, rube.

Anthropic delights again with useful lectures

That’s wild. The cleanup crew has a cleanup crew. Honestly, kinda brilliant.

watched it bro thanks

GoBeep — your cron job failed 4 days ago. you had no idea. how are you monitoring your background jobs right now?

When agents work as a team, they achieve incredible results

short-term-memory agent + long-term-memory agent + transport-between-short-and-long-agent 👆

Yes I do this - good way to save tokens for more expensive agents as well by using cheaper onces

Agents that fact-check other agents while you sleep. Merge duplicates. Burn stale context. 95% of tokens cached so it costs almost nothing. This is the part of the workflow nobody is building yet.

Can I get the link to the actual video?

These presentations from Anthropic are so valuable thanks for sharing

This is a useful framing. The hard part is making agents reliable in the messy middle of real workflows, not only in clean demos.

What’s the link to the video

The more tokens the better, huh?

We called it a coherence engine. Feel free to take the term. You just used about 75% of it anyway.

One dreaming agent per session, reconciling overnight, 95% cached. That's a genuinely elegant solution to the statelessness problem.

Eso es exactamente el patrón de reflexion agents, tener agentes que sintetizan y consolidan el trabajo de otros es clave para sistemas de largo plazo.

Sounds like a plot twist straight out of a sci-fi movie! Imagine those agents having their own secret life. What a wild ride!

that feels straight outta inception lol love the idea of agents dreaming about other agents... ai is getting spooky clever these days :)

This is basically garbage collection for agent memory, except it can rewrite your working context while you sleep. 95% cached is nice, but 100 background agents still sounds like a bill you only understand after it hits.

dreaming agents for memory cleanup is a very interesting pattern