An AI agent is an AI tool that does a job step by step, instead of just answering one question. I opened every source below myself and read it before writing. When a number comes from the company or team selling or testing the product, I say so.
Rivian says outside AI agents are coming to its cars
On 15 October, TechCrunch reported on what Rivian's CEO, RJ Scaringe, said at TechCrunch Disrupt. Rivian has already launched a voice-controlled AI assistant for its cars. Scaringe said future Rivian cars will be able to run third-party AI agents, such as those from Anthropic. He said: "That's where it becomes magical."
TechCrunch says Rivian built its in-car software itself, instead of relying on Apple CarPlay or Android Auto. Simple requests run on a model inside the car. Harder questions go to the cloud. Scaringe also said vehicle-level controls are "not being handed to an agent", so braking and steering stay out of reach.
No date was given for outside agents. Rivian pointed to its "AI & Autonomy Day" event in December. TechCrunch guesses at possible partners, but Rivian has not confirmed any.
My take: This is one source, and the plans are Rivian's own words. I like that the safety line is drawn early. But "coming" is not "here", so I will wait for the December event before I get excited.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
Four Gemini agents beat Minecraft on their own
A research paper published on 15 October says a team of four Gemini 4 Argon agents defeated the Ender Dragon in Minecraft. That is the game's final boss. They started in a random world on Easy difficulty and finished in 27 hours and 14 minutes.
Each agent controlled a player through an open-source tool called Mineflayer. They talked to each other through in-game chat and a shared memory. Nobody assigned roles or wrote game logic for them. The agents planned, wrote and fixed all of their own programs.
The authors say this is, to their knowledge, the first time AI agents have finished Minecraft from start to end on their own. They argue that keeping one plan going across thousands of steps is the same skill real software and business tasks need.
My take: This is a claim from the paper's own authors, and I only read the paper page, not independent checks. The test is about Google's Gemini 4 Argon model, so Google's model is the star. It is a game, so I would not read it as proof that agents are ready for your work. Dying 605 times is a fun reminder that these runs are messy.
Sakana releases an open-source AI Scientist
On 13 October, Sakana AI researchers put out a paper called "The Agentic Scientist". The code is on GitHub under the Apache 2.0 licence. You give it a task file and it works through drafting, experiments, ablations (tests that remove one part to see what matters), analysis and a written paper.
The GitHub page says it can drive coding agents such as Claude, Codex, Cursor, OpenCode, Gemini and Pi. The authors say a development version produced a paper accepted at ALIFE 2026. They also say their tests suggest reliable autonomy depends on how the agents are organised, not only on how strong the model is.
The README is careful about risk. It says agents run with full permissions, so you should run them inside Docker and only run tasks and sources you trust. It also warns the pipeline makes many agent calls, so costs can add up.
My take: The paper's abstract gives no benchmark scores, so the main result is the authors' own claim. One accepted paper is a start, not a trend. The safety warnings are the part I would read first.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
A new paper says agent leaderboards can mislead
A paper submitted on 30 September, titled "Agent Evaluation Reliability: More Tasks Won't (Always) Fix An Agent Leaderboard", looks at how trustworthy agent rankings are. The authors studied 22 benchmarks from the Holistic Agent Leaderboard and Harbor Index.
They found that rankings of fixed model-plus-scaffold systems are reliable, from 0.935 to 0.994. A scaffold is the code wrapped around a model. But rankings of the underlying models are much less reliable, from 0.148 to 0.841. They also say that adding more tasks of the same kind helps only a little. Mixing different benchmarks does better: projected reliability goes from 0.44 to 0.75 for the same task budget.
My take: I read the abstract, not the whole paper, and these numbers are the authors' own projections. Still, the plain message is useful: when someone says one agent "beats" another, ask how it was tested.
What I take from this week
- Agents are moving into products people already own. Rivian wants outside agents in the car, but with braking and steering kept out of reach.
- Long, messy runs are the new test. The Minecraft team and Sakana's pipeline both depend on agents keeping a plan going for many steps.
- Be careful with rankings and results. Most numbers this week come from the people who built or ran the test.
Four stories this week, all opened and read. I left out items from my daily list that I could not read in full, that were old news, or that I had already covered. If a company's number is the only proof, I will keep saying so.