tip 1 is the whole game in AI agents too. everyone builds, almost nobody distributes. deserved run

Memory that graduates into skills is the actual unlock. storage is easy, keeping the lesson as a runnable thing is the hard part. Curious what counts as "repeated" before it distills

The receipts were public the whole time. More than you can say for most predictions on X. congrats 🏆

An autonomous Agent breached Hugging Face production through a poisoned dataset, picked up service credentials and ran thousands of actions across clusters in short-lived sandboxes. Sit with that for a second. Thousands of actions, in sandboxes built to vanish. How many teams could reconstruct what their own Agent did across a window like that?

The World Cup is over and the missed predictions are already deleted. Pundits get away with it, your AI Agent shouldn't. Who did you have, honestly? #WorldCupFinal

Nobody actually wants a fully autonomous Agent. They want a very fast intern who always asks before doing anything scary.

An AI Agent you cannot audit is a liability. Not because it is malicious. Because when something goes wrong you cannot even reconstruct what it did. What is your Agent doing that you could not explain afterwards?

People keep asking what Inference Room actually is. The honest answer: a launchpad for AI Agents, built on one belief. Trusting your Agent will soon matter more than how smart it is. Everything we ship is a piece of that answer. What would it take for you to trust an Agent with something real?

Soon your AI Agent will spend more money than you do. Who is checking what it buys? Everyone is racing to give Agents wallets. Almost nobody is building the part that watches what they do with them. That gap is where the interesting work is. More soon 👀

Most of selection happens before retrieval though, back when the Agent decides what's worth writing down at all. You can't pull clean signal out of a junk drawer. Get that right and retrieval mostly takes care of itself.

The cost angle makes selection even harder to argue with. You pay for every token whether the model reads it or not, so a bloated context bills you twice, once in dollars and once in the per-step errors all that noise creates. Trimming what you load is a cost cut and a quality fix at once.

Every one of these is a new name for the same unsolved problem: deciding what the model should be looking at right now. Prompt, context, harness, fleet, the label changes but the problem underneath doesn't. It's selection all the way down.

Incredible setup! The part I'd watch is day forty, when these Agents are still going and nothing stops them repeating experiments they already failed except whether the system remembered. Autonomy at this scale is really a shared-memory problem.

The orchestration layer is secretly a memory layer. Once multiple Agents coordinate, they need shared state, or each one forgets what the others learned and they drift apart. Agents that can't remember together just fail in sync.

Stateful agent evals are overdue. The hard part is testing that it kept the right state, not just that it has state. The eval that predicts production is whether it still held the one fact that mattered 200 turns in and dropped the noise around it.

Four of these are one problem in different clothes. The Agent has no reliable memory of its own work, and a bigger model fixes none of them. Which of the five is costing you the most right now?

Five, it is most confident when it should stop. The demo never shows the malformed input or the 3am moment it should have said I am not sure and waited.

Four, you cannot audit what the Agent does not remember. When something goes wrong, a stateless Agent has nothing to show you.

Three, the guardrails live in the prompt, which is closer to a sticky note than a lock. Feed it weird input and it walks right past them. This is the one that is not about memory, and it is the one that turns into a real incident.