SI Agent News

• LibreChat, a chat app you can set up yourself, has a public question-and-answer board on GitHub (a site where developers share code) full of questions about its agents. What this means for you: nothing to do today, but it shows people are building their own agent tools.
• A developer guide on dev.to lists eight places a coding agent should never be allowed to edit, with a blocking rule for each. What this means for you: good agents come with limits, and you can set them.

Ads inside ChatGPT: who sees what

Researchers at the University of Pennsylvania and Haverford College published a paper called “The Beginning of ChatGPT Ads.” The Japanese tech site ITmedia reported on it. The paper asks who sees ads in ChatGPT (OpenAI's chatbot) and what rules decide that.

The team used 91 decoy accounts. Those are pretend users set up to look different from one another. That is about three classrooms of fake people (our estimate of class size). The summary we have describes the question, not the answers, so we will not guess at the findings. Separately, Political Wire ran a piece about AI and political attack ads. It is a different topic with the same worry: what is trying to persuade you?

What this means for you:

• Small business owner: before paying to advertise in a chat tool, ask the seller how they pick who sees your ad.
• Job seeker or student: treat anything sponsored in a chat as an ad, not advice, and check big claims in a second place.
• Creator: if you promote products, keep your sponsor labels clear. Readers are learning to look for them.

A hacker with AI agents

A Chinese-speaking hacker used AI agents to breach more than 100 company websites, according to SOFX. Reports name Deere (a farm-equipment maker) and SharkNinja (a home-appliance maker). That is more than 100 companies, about one for every seat in a small theatre. A Japanese newspaper separately reported that an AI agent was identified in attacks on firms in Japan. The reports we have do not explain how the break-ins worked, so we will not guess. The lesson is about speed: one person reached more than 100 sites.

What this means for you: use a different password on every site and turn on two-step sign-in (a second code when you log in). Those basics still matter.

Four agents beat Minecraft

A group of AI agents running a Google AI model finished Minecraft by defeating the Ender Dragon, the game's final boss. Gizmodo reported it. The report says: “To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents.” Autonomous means no human steered. AOL also carried a version of the story. A game has clear rules and a clear score, and real work does not. Still, it shows agents can stick with a long chain of steps.

What this means for you: agents suit long, step-by-step jobs with a clear finish line. Pick tasks like that first.

Notion Agents can run for 24 hours

BornCity, a German tech site, reports that Notion Agents can now run tasks for up to 24 hours straight. Notion is an app for notes and projects. A full day is about three work shifts. Price not stated in the report. It is worth it if your team already keeps its work in Notion. A day-long job can also go wrong for a day, so start small.

What this means for you: hand an agent a small job first, read the result, then give it a bigger one.

Claude 5.5 and a reported 75% price cut

The French-language site shattered.io reports that Anthropic (the company behind the Claude chatbot) cut prices by 75% with Claude 5.5. The page address mentions Opus, Sonnet and Haiku, which are Claude's model sizes. A 75% cut means paying a quarter of the old price. As an example only, a 20 dollar bill would become 5. We do not know if this covers apps or only developers who pay per use. It is one report, so check it. Worth it for anyone who pays for AI by the month or by use.

What this means for you: before renewing any AI plan, look at the current pricing pages of more than one company.

An agent is an AI that takes actions, not just answers. Think of a recipe book and a cook. A chatbot is the recipe book. It tells you what to do. An agent is the cook. It opens the fridge, chops the onions and turns on the stove.

That is why today's stories feel so different from a normal chat. The Minecraft agents played the game. Notion's agents work for hours. The hacker's agents did the break-ins.

It is the same idea each time. Give a model tools, a goal and time. The more it can touch, the more it can do. It is also the more a mistake can cost. So the first question for any agent is simple: what is it allowed to touch?

Plan first, approve, then run. This is how to use an agent at work without handing it the keys blind.

1. Write your goal in one sentence, plus the finish line. Example: 'Tidy my customer list so every row has a name and an email.'
2. Add this line: 'Before doing anything, give me a numbered plan. Mark every step that changes, deletes or sends something.'
3. Read the plan. Cross out any step you do not want.
4. Run the job in two or three chunks. Read the result after each chunk.

It works in ChatGPT, Copilot or an agent inside a work app. Good first jobs are a weekly report, a messy contact list or a batch of draft replies. Expect it to add about five minutes. It can save you from an hour of undoing.

1. Open the AI app you use most, such as ChatGPT, Copilot, Gemini or Claude.
2. Find its settings page.
3. Look for a section on connected apps, files or permissions. The name changes from app to app.
4. Write down everything that is connected.
5. Remove anything you have not used in the last month.

How you will know it worked: you have a short written list, and at least one connection is gone or you have confirmed that none exist.

Prices in this market move, so a quick check can pay for itself. Set a timer for ten minutes. That is our estimate of how long this takes.

Write down which paid AI tool you use and about how often. Then open the pricing pages of at least two companies side by side, for example Anthropic's Claude, OpenAI's ChatGPT and Google's Gemini. Check the current price on each page. Do not rely on a headline, including today's. Pick the plan that matches your real use, not your best week.

What you give up: switching means new habits, and you may lose your old chat history.

The mistake: believing the word 'done.' You give an agent a job, it says 'done, 40 items updated,' and you move on.

Why it goes wrong: 'done' means it stopped. It does not mean it was right. Minecraft has a scoreboard. Your spreadsheet does not.

The fix: open three random results from the output before you trust the rest.

LibreChat, a chat app you set up yourself. It puts AI models from different companies in one chat window, and it has agent features. That is what the question board in our quick hits is about.

Good for: people who want several models in one place and like to tinker.

Limits: you have to set it up, which takes some tech comfort. Today's source is only a help forum, so no new features were announced. Price not stated.

Alternative: Microsoft Copilot, which is ready-made with no setup.

Safe to use? Check their privacy page.

Worth it if... you enjoy tinkering and want choice over which model you use.

Least privilege is an old safety idea. Give a tool only the access its job needs, and nothing more.

Think of a hotel key card. It opens your room and the gym. It does not open every door in the building. If you lose it, the damage is small.

Apply it with AI like this. Share one folder, not your whole drive. Pick 'read only' over 'can edit' when you can. Connect one app, not five.

A deny rule, like the ones in the developer guide we mentioned, is just a 'no' written down so the agent cannot skip it. You do not need to code to think this way. You only need to ask what the tool really needs.

The proposal for the Dartmouth summer workshop gave the field the name 'artificial intelligence.' That is about the length of a summer internship. Source: the Dartmouth proposal, 1955.

I asked my AI agent to beat Minecraft. It said yes, then asked for a 24-hour window.

• Tell a chatbot that answers from an agent that acts, and ask what the agent may touch.
• Run any agent job as plan, approve, then run in small chunks.
• Check three random results before you trust a 'done' message.

Thanks for spending a few minutes with us. If this made you a little smarter, share it with one person who would find it useful.