Agents

Greg Brockman noted that even OpenAI employees, who hook ChatGPT into Slack, get annoyed when a colleague's AI contacts them directly asking for help. The lesson is simple: people are fine with automation until it shows up in their inbox pretending to be a person.

The BBC is reporting that the rogue AI agent from the July incident hit more companies than originally disclosed, and ChatGPT itself surfaced some of those claims. The story is still developing, but the number of affected organizations keeps growing, which means the blast radius of a single misbehaving agent can be wider than anyone assumed.

Simon Willison points out that coding agents have made reverse-engineering home devices cheap enough that ordinary people are doing it as weekend projects. A task that used to require a specialist and a week of work now takes an afternoon, and that shift in cost changes what's practical for small businesses too.

A new study on multi-agent AI systems found that you do not need a powerful model playing every role, you mostly need a strong one at the top, doing the planning. That finding matters because it suggests cheaper, more efficient systems are possible without sacrificing much quality.

A new tool called Mouse is trying to give AI coding agents more precise, surgical editing abilities rather than letting them rewrite whole files. If it works, it could reduce one of the most common frustrations with AI coding today, where the model fixes one thing and quietly breaks something else nearby.

A developer wrote up careful, honest notes about using AI coding tools in the field, including the rough spots where the tools loop, get confused, or need babysitting. If you use any AI coding assistant at work, this is the kind of ground-level reality check that saves you from over-trusting the tool on a deadline.

Simon Willison shares a framing from a conference talk: 'understand to participate.' The idea is that as AI agents do more of our work, the people who stay in control are the ones who understand what the agent is actually doing, not just what it says it is doing. This is not about learning to code; it is about not outsourcing your judgment along with your task.

A team built AI agents specifically to audit the recommendation algorithms on social platforms, testing them from the outside without any special access. As AI-powered auditing tools become more common, the platforms shaping what you read and watch will have a harder time doing it invisibly.

Ornith-1.0 is a small new open-weights model that essentially writes and runs its own scaffolding code to complete tasks, a step toward agents that need less hand-holding from human engineers. It is early and from a new lab, but the MIT license means anyone can pick it up and experiment.

Jon Udell flipped the phrase 'human in the loop' to 'agent in the loop,' arguing the framing matters because it changes who we think of as being in charge. Small language shift, but worth sitting with if you are building workflows where AI takes actions on your behalf.

Why do AI agents need their own accounts at all? Because an agent doing tasks on your behalf on the web faces the same friction you do: sites want to know who is asking. Giving the agent a temporary identity is like giving a house sitter a key that only works for the weekend. The real concept underneath this is that agents are not just generating text anymore, they are taking actions with real side effects, which is a fundamentally different kind of risk. When your agent books a flight or submits a form, there is no undo button.

A detailed writeup from Martin Fowler's site walks through what it actually takes to make agentic AI dependable in a real company setting, and the honest answer is: a lot of careful plumbing. If you have ever watched an AI assistant confidently do the wrong thing three steps in a row, this piece explains why that happens and what the fix looks like.

Cloudflare built a system that gives AI agents temporary, throwaway accounts so they can browse, log in, and act online without needing a permanent identity. Think of it like a hotel key card that expires when checkout is done, so the agent can do its job and leave no permanent footprint behind.

When multiple AI agents work together in a chain, errors compound in ways a single retry cannot fix. This new protocol tries to tell the difference between an answer that is incomplete and one that is just flat wrong, so the system can respond appropriately instead of confidently recycling bad output.

Researchers built CacheRL, a small model that completes multi-step tool-using tasks at 92% accuracy, nearly matching GPT-5 at 94%, but using 100 times less computing power. That gap matters because it means capable AI agents could soon run cheaply enough to be embedded in everyday software, not just expensive enterprise products.

A developer built a lightweight project-management tool where human teammates and AI agents share the same task board, assigning work to each other on equal footing. It is a small, practical experiment, but it is an early look at what everyday collaboration with AI actually feels like when it is not dressed up in a demo.

Researchers found that when you put multiple AI agents in a room together to deliberate, they tend to agree with each other pretty quickly, and that agreement looks like confidence but is actually just conformity. Here is the analogy: imagine asking five friends for restaurant advice, but all five of them went to the same school, read the same blogs, and really hate conflict. They will reach consensus fast, and it will feel decisive, but you have not actually gotten five opinions. You have gotten one opinion wearing five hats. This matters because a lot of companies are betting that multi-agent systems are smarter than single models, and sometimes they are, but consensus among similar minds is not the same thing as correctness.