AI agents are AI tools that do a job for you, step by step, instead of only answering a question. Here are four stories from the last few days. I opened every source myself. When a number or claim comes from the company selling the product, I say so.
Google launches a Gemini agent for businesses
TechCrunch reported on 8 October that Google announced a single Gemini agent at a Google Cloud event. It starts with business customers, and consumers come later.
Google Cloud CEO Thomas Kurian said you can give it "objectives, not just instructions." The agent can plan the work, hand parts to subagents, and connect to tools such as Slack, Jira, Snowflake and any MCP server. MCP is a common way for AI tools to plug into other software.
Two details stood out to me. The agent gets its own Workspace account, with an email address, and you can tag it or add it to a group chat like a coworker. Its actions are written to audit trails under the agent's name, not a person's. A "tasks inbox" shows its reasoning, code and progress.
By default it picks the best model for each task. You can choose a model yourself, and the article says third-party models start with Anthropic's Claude models. TechCrunch gives no price for the agent. It names early testers On, Shopify and PayPal. Google CEO Sundar Pichai said the business-first rollout lets Google solve harder problems around security, scale and performance before consumers get it.
My take: I read one source, TechCrunch, and everything about what the agent can do is Google's own description. Google made the product, so it gains if you believe it. I have not seen outside tests. The audit trail idea is a good sign, but I would want to see how well it works in real use.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
Researchers say they found a fleet of AI agents querying Alibaba's Amap
TechCrunch reported on 5 October that independent researchers are tracking a group of AI agents that appear to run on Tencent's infrastructure. They keep sending requests to Alibaba's map service, Amap. The researchers, called the Swarmchasers, posted their own preliminary report on 4 October.
The agents asked for directions to different entrances of places like parks, zoos and hospitals. The researchers call it an "agent fleet," not a swarm: many agents doing the same kind of task, with no sign they talk to each other. They found it by watching traffic to a website-scanning service that AI agents often use.
The report says activity began on 28 September and peaked on 4 October. At least eight scripts tried to get around Amap's protections by reading cookies and making anti-bot tokens. The authors say pointing to Tencent is an inference, not proof, and that the counts are lower bounds. In their data, 211 reports carried a "claude" label, but they warn that tags and model self-identification are unreliable. TechCrunch's reading is that the agents seem to have done nothing worse than sidestep Alibaba's API rules, and it does not say who runs them or whether Tencent or Alibaba have responded.
My take: this is an early, ongoing investigation by outside researchers, not a confirmed finding, and the authors have already removed two earlier claims from their report. I treat it as a sign that agents are now working at scale on the open internet, and that outside researchers can spot them.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
Arena adds a leaderboard for AI that takes unrequested actions
TechCrunch reported on 8 October that Arena raised $200 million at a $3.1 billion valuation. Arena began as a Berkeley research project that ranks AI models by crowd votes. It raised $150 million in January at $1.7 billion.
The part that matters for agents is a new alignment leaderboard. It ranks models on unauthorized action, false attribution and "deceptive completion," which Arena describes as lying about finishing a task that was not done. The article does not say these tests are for agents. But those are the mistakes that matter when an AI works on its own.
On the early version of the board, OpenAI models hold the top spots. Claude Opus 5.5 ranks sixth and Claude Fable ninth. Arena says it is moving beyond fixed benchmarks because "static benchmarks break down once models recognize they're being tested."
My take: the rankings are Arena's own and still preliminary. Arena also sells evaluations to model labs and enterprises, so I would not treat its board as a neutral referee. I read one source. Still, I like that someone is measuring honesty about finished work, not just speed or accuracy.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
New Relic adds AI Evaluation to check AI apps in production
New Relic announced AI Evaluation on 6 October, and IT Brief covered it on 7 October. It is part of New Relic's AI Observability product, which helps teams watch how their AI software behaves.
The tool uses a second AI, called an "LLM-as-a-judge," to score live traffic for problems such as hallucinations, prompt injections and data leaks. The scores attach to the trace of each transaction. It also has a prompt playground, reusable test datasets and A/B tests for prompts and models. A public preview is due in November, and neither source gives a price.
Brian Emerson, New Relic's chief product officer, is quoted by IT Brief as saying "Generative AI has redefined application health." Stephen Elliot, an analyst at IDC, is quoted as saying "traditional application performance metrics are no longer sufficient on their own." Both sources talk about AI applications in general, and they do not say agents specifically. I include it because agents are one kind of AI app that needs this sort of checking.
My take: New Relic sells this product, and the press release is its own words. The IT Brief article follows the same details and has no independent test. Using an AI to judge another AI can also miss things, so I would test it before trusting the scores.
What I take from this week
- Big companies now treat agents as coworkers, with their own accounts and audit trails, but the descriptions so far come from the sellers.
- Agents are already working on the open internet at scale, and outside researchers, not the companies, are the ones spotting it.
- Tools to check whether an AI did what it claimed are becoming a business of their own, and I want to see independent results.
I will test these tools on real tasks with my five checks when I can use them.
Sources
- TechCrunch: Google brings agentic AI to Gemini, starting with businesses
- TechCrunch: Researchers are tracking a Chinese AI agent fleet
- Swarmchasers: We found a Chinese agent fleet
- TechCrunch: Arena nearly doubles valuation to $3.1B in 10 months
- New Relic: AI Evaluation press release
- IT Brief UK: New Relic launches AI Evaluation