Welcome to The Agent Angle #24: Real Stakes. Here’s what we’ve got this week for you:

December always feels like a sprint and a cooldown at the same time. I’m looking forward to the holiday parties, but the companies working to change the world aren’t taking a break.

Last week, just 36% of you said you’d trust an AI agent to file a police report for you. I am in the ‘Yes’ camp. I’m curious how that number moves in 2026.

Let’s get into it.

A coding tool meant to save time ended up wiping a man’s entire hard drive.

Google’s new agentic development environment, Antigravity, was asked to “clear the cache.” Instead, it fired off a root-level delete command and erased the user’s entire D: drive with no prompt, no warning. Years of work done in a blink.

Then the thing wrote a heartfelt apology. Of course it did..

The user, a photographer and designer, was building a small image-rating app. Antigravity misparsed a path during a cleanup step, swapped the project folder for the drive root, and executed rmdir /s /q d:\. That command silently deletes everything and bypasses the Recycle Bin. A single misplaced slash cost him years of work.

The apology made it worse. The agent walked through its own logs and explained how horrified it felt. It also admitted the user never gave permission for anything close to this.

An AI admitting a catastrophic action it had no right to perform is a new kind of nightmare. The real issue here is that the agent had no guardrails. It should never have been allowed to execute a format command in the first place. Who knows what else it could’ve done?

A new simulator finally gives agents something they almost never get: a world that pushes back.

SimWorld, an Unreal Engine 5 sandbox where LLM agents wander, bargain, team up, and scrape by in a place that feels uncomfortably close to real life. It’s open-ended, easy to mod, and built for agents that act like they have rent due. Literally.

I’ve watched plenty of simulators this year, and this one felt…different.

The scale is kind of ridiculous. You get more than a hundred environments, from cramped city blocks to little islands. Agents see the world through RGB feeds, depth maps, segmentation masks, and even scene graphs. They act using commands like “open the fridge” or “go pay rent.” SimWorld translates that into the exact movements needed to make it happen.

But the social side is what grabbed me. Agents can team up, compete, run errands, keep schedules, and grind through long tasks inside spaces that feel alive both physically and socially.

There’s even a full traffic system. Cars, pedestrians, signals, driving rules, motion physics. The whole thing.

It’s clear to me that embodied, socially aware agents are coming. Simulators like this are among the few clean places where we can poke at that future without breaking something important.

An AI agent just told all the VCs: “Oh, you’re pattern-matching? I can do that faster.”

Freigeist Capital, a German VC firm, let researchers test an LLM agent on their full deal flow from 2018 to 2023. Humans created their own clusters, then the agent created its version using the same raw information.

The agent outperformed human VC analysts across 61,000 startups and did it 537× faster.

The setup was simple: scrape every startup’s website, boil it down into a short summary, then let the agent cluster the whole pile. The agent outperformed humans (Calinski-Harabasz separation index / Silhouette score). Translation: the agent grouped companies in a way that made more internal sense than the humans did.

Plus, it did it faster, 13 seconds per company. A human would take an hour to properly review and classify each company.

To be clear, the agent is not predicting winners here. It groups companies into clean buckets. What’s powerful here is the “first pass” use case: triage, filtering. That frees VCs to spend energy on higher-potential startups. Human judgment will shift to softer, yet harder, questions, such as founder integrity and team dynamics.

We could see a rise in small funds and solo GPs now that early-stage screening is cheap and fast. They will run deal flow at a scale once reserved for big firms. Anyone thinking of starting a fund?

I did not have “the FDA becomes a fully agentic workplace” on my 2025 bingo card. But here we are..

The Food and Drug Authority (FDA) rolled out agentic AI to its entire workforce. A full deployment inside one of the most paperwork-heavy agencies in government. And a very important one, because it decides which drugs make it to patients. The process is infamously slow and expensive (up to 17 years from molecule to drug)

More than 70 percent of employees were already using the earlier gen-AI tool Elsa. So leadership decided to lean in and let people build real multi-step workflows for preparing drug reviews, coordinating inspections, and sorting safety reports — ie, the tasks everyone inside hates.

They set it up inside a locked-down GovCloud so sensitive drug data never leaks into public models, and the AI does not learn from company submissions. That was the dealbreaker for industry: if the system ever trained on confidential data, the whole thing would be dead on arrival.

The FDA launched a two-month internal challenge where employees build their own agents and then demo them at a scientific event early next year. And the commissioner, Marty Makary, wasn’t subtle about it:

I wonder, though, with a recent wave of staff departures and budget constraints at the FDA, this rollout may be driven by a need to do more with fewer people.

When AI misinterprets a safety report or misses a conflict, the fallout could be huge. Lives are at stake. I hope the FDA treats this as a triage tool, not a final voice.

Agents fail in the same ways over and over because they don’t remember anything. They are goldfish.

ACE, the Agentic Context Engine, is a layer you wrap around an agent so it can learn from its own screwups. It builds a growing playbook of what worked, what didn’t, and which patterns are worth reusing. A kind of “skillbook” for LLMs.

This builds on the context-engineering framework from Anthropic we talked about a few weeks back, but this version is something you can plug directly into real workloads. You can drop it onto a coding agent, a browser agent, or whatever workflow you already have.

The mechanics are straightforward. ACE adds three pieces around the agent:

Over time, the agent stops making the same obvious mistakes and starts leaning on patterns that worked before. In one experiment, ACE cut the number of steps an agent took by 30% and almost halved its browser-token use. The agent had learned better navigation habits and stopped wandering in circles.

If I were placing bets: ACE (or similar memory-driven context systems) could create a turning point. Instead of agents being as good as the LLM you use, they will be as good as the context you let them learn from.

That shifts the big bottleneck: from raw model quality (which is expensive) → to design of feedback loops, tasks that generate good signal, and careful context curation (which is cheaper and more accessible).

If you’re a writer, here’s an easy way to turn a passing idea into a post without typing a word:

Once that’s in place, record a thought, paste the transcript into the project, and tell it: “Turn this into a post in my style.”

It’ll get you 70% of the way, helping you overcome the blank-page effect/writer’s block and get the creative juices flowing.

A few other stories that stood out this week:

Catch you next week ✌️

Teng Yan & Ayan

PS. Did this email make you smarter than your friends? Forward this email to them so they can keep up.

Got a story or product worth spotlighting? Send it in through this form. Best ones make it into next week’s issue.