gmo retweeted
In town for Config? Come work from the Browserbase office ☕️
Fast Wi-Fi, free coffee + matcha, a quiet place to focus, and a chance to meet designers + the Browserbase team.
Finished my first read of "AI Control: Improving Safety Despite Intentional Subversion" last night.
Such an underrated and fascinating corner of this space.
premise: Assume the model is adversarial, design the environment so it can't cause harm anyway.
I have been building sandboxes, honeypots, and simulation and real environments for threat actors to "explore" for a long time.
so I am hooked.
Shout out to @redwood_ai for the work here. arxiv.org/pdf/2312.06942
K8s Operators were agents before agents were cool.
Observe state → compare to desired state → take corrective action. That's operational knowledge encoded directly into software, running in a loop, 24/7.
If you're building agentic systems and haven't read up on them, I highly recommend it.
kubernetes.io/docs/concepts/…
I just stumbled upon this article
On MCP Real-Time Streaming
Highly recommend
chatforest.com/guides/mcp-re…
A lot of my pains in agent harness design are not memory problems. They are state divergence problems.
The agent remembers what it believed.
The problem is that reality changed.
A service became unavailable.
A capability was added.
A tool changed.
Infrastructure evolved.
An exploitation attempt already succeeded or failed.
The agent continues reasoning from stale state and arrives at the wrong conclusion.
The more I study probabilistic robotics, the more I realize many of the problems I’ve been wrestling with look less like memory problems and more like state estimation problems.
Spent most of today thinking and researching about agent architecture.
One thing I gained a deeper appreciation for is autonomous vehicle systems.
I started looking at memory, capabilities, events, and planning. The further I dug, the more I realized autonomous systems start somewhere different:
They start with a model of the world.
Not memory.
Not tools.
Reality first.
Everything else builds on top of that.
Alright, let me go get smart...
arxiv.org/abs/2502.16090
I'll be right back
This is what I mean when I talk about memory infused with time.
An LLM can remember that we discussed something before, but it doesn't naturally know whether that conversation happened 5 minutes ago or 5 days ago. It has memory, but not temporal context.
Without time, an agent cannot reliably reason about freshness, decay, or whether reality may have changed since the memory was created.
Yesterday GPT and I worked on a project all day. We finished, I slept, came back the next morning, and asked to start Phase 2.
GPT told me we should rest because we'd already done enough for the night.
The interesting part is that it remembered the project perfectly.
What it didn't understand was that "last night" had become "this morning."
The memory was accurate. The temporal reasoning was not.
Agent memory needs more than facts. It needs time.
One problem I haven't heard discussed much in agent engineering:
How does an agent learn that reality changed?
A new capability is added.
A provider becomes available.
A tool gains new functionality.
Infrastructure evolves.
I don't want to keep reloading all of this into the context window.
Git knows what changed.
The agent doesn't.
I'm starting to wonder if long-lived agents need something closer to an operational bulletin/event stream where they can simply ask:
"What has changed since I last looked?"
Curious if others have run into this problem - or if I am just going crazy.
for example:
We originally had a post_config deployment entrypoint. Later I added a pre_config entrypoint.
The agent didn't know this new capability existed, so it spent time looking for alternative implementations and workarounds. It wasn't until I forced it to reread the code/docs that it discovered the new capability.
This makes me wonder if agents should be able to ask:
"What capabilities have been added to deployments since I last looked?"
and receive capability registrations linked to the corresponding Git references.
Even better, perhaps these capability registrations should be emitted as events so interested agents can update their world model when the change happens rather than discovering it later through retrieval.
Most agent memory discussions focus on storing facts.
I'm starting to think a missing piece is time.
Not just what happened.
But:
How long has this been true?
How long since we checked?
How long have we been investigating?
Engineers use time as evidence. Most agents don't.
