SI Agent News

• PolyAI, which builds AI that talks with customers, launched a Builder MCP. MCP (a shared plug that lets AI tools connect to other software) lets developers build PolyAI agents without leaving their coding program. What this means for you: if you never build agents, skip it.
• A PolyAI profile follows a linguist who teaches business AI how people really talk. What this means for you: next time an AI phone line answers, notice the pauses and wording. Someone designed them.
• PolyAI explains its ADK, a toolkit for building and launching dialog agents (AI that holds a back-and-forth conversation), so any team can use it. Price not stated.

1. Claude Duet, an always-on agent from Anthropic

Anthropic, the company behind the Claude AI assistant, has announced Claude Duet. It is being described as a 'proactive agent.' An agent is an AI that takes actions, not just answers questions. Duet can run in the background, surface tasks to you, and handle a variety of work.

Most chatbots wait for you to type. Duet is built to speak up first. Coverage frames it as Anthropic's answer to Meta and OpenAI. Price: not stated. An always-on tool raises a practical question: what is it allowed to see and do while you are not watching? Check that before you switch anything on.

What this means for you:

• Small business owner: a background helper could flag follow-ups you would miss. Start by letting it watch one task.
• Job seeker or student: learn to give an agent clear rules now. That skill carries over to every assistant.
• Creator: hand off dull chores like tracking replies and deadlines. Keep the creative work yourself.

2. AI finds a flaw that took websites offline

Reports say AI found a 'critical' security flaw in PHP, a programming language that runs many websites. Cloudflare, a company that helps protect websites, said its chief executive saw a dramatic surge in attempts to exploit the flaw once it was announced. To exploit a flaw means to use it to break in.

AI was part of how the problem was found, so these tools are now working on the security side. And the speed of the attempts shows how fast an announcement can turn into attacks. The danger window is the time between 'announced' and 'fixed.'

What this means for you: if a site you use is down, this may be the reason, not your device. If you run a site, ask your host whether the fix is in.

3. ChatGPT moves into Chrome

TechRadar reports that ChatGPT, OpenAI's chatbot, now has an integration with Chrome, Google's web browser. The writer tested whether it really makes browsing better. The headline says Gemini, Google's own assistant, now has serious competition in its home browser. The summary does not say exactly what ChatGPT can do there, and the price is not stated.

The browser is where most people spend their working day, so whichever assistant sits closest to the page has an edge. Competition inside one browser tends to push both to improve, and you now have a real choice about which assistant sees the pages you read.

What this means for you: if you already use ChatGPT, it may now be one click away. If not, nothing here is urgent. Read the hands-on test first.

4. Chrome adds Gemini scam warnings

CyberInsider reports that Chrome is adding Gemini-powered warnings for ClickFix and other scams. ClickFix is a trick where a fake page says something is broken, then tells you to copy and paste a 'fix' that is actually harmful.

This is one of the more practically useful stories today. These scams work because they make you do the dangerous step by hand, so a warning at the right moment can stop the damage. The coverage does not say how well the detection works, so treat it as an extra layer, not a guarantee.

What this means for you: a warning is a safety net, not a shield. No real website needs you to paste commands into your computer. If a page asks, close it.

5. Walmart lets AI agents shop for you

The Indian Express reports that Walmart launched tools so AI agents can shop for customers. An agent could, in principle, do a shopping run on your behalf. Fees and details are not stated.

Walmart is making its store usable by agents, not only by people with a browser. If that catches on, the question shifts from 'where do I click?' to 'what did I tell my agent to buy?' Your instructions, including budget, brand and substitutions, become what matters.

What this means for you: set a spending limit and check the cart before you pay. Start with a low-stakes list, like household basics.

6. PolyAI's Builder MCP brings agent building into your editor

PolyAI has introduced the Builder MCP, which lets people build PolyAI agents without leaving their editor, the program where developers write code. MCP is a standard way for AI tools to connect to other software. The summary gives no more detail on features, and pricing is not stated.

The idea is convenience: the work happens in one place instead of across a code editor and a separate building tool. Fewer tool switches usually means faster tries and fewer small mistakes. It also signals that agent building is moving into everyday developer tools.

What this means for you: if you or your team write code, read PolyAI's announcement to see whether it fits. If not, skip it for now.

7. The linguist teaching enterprise AI how people talk

PolyAI also published a profile of a linguist, an expert in how language works, who teaches enterprise AI how people actually talk. The summary gives no more than that headline, so the specifics are in the original piece.

Real speech is messy. People interrupt, hesitate and change their mind mid-sentence. An AI that handles customer calls has to cope with that, not just tidy written sentences. A language expert on the problem suggests good voice AI depends on understanding people, not only on bigger models.

What this means for you: test a voice or chat assistant with the way you really speak, not perfect phrases. That is the fairest check.

8. ADK explained: how PolyAI lets any team build dialog agents

A third PolyAI piece explains its ADK and how it lets any team build and deploy dialog agents. A dialog agent is an AI built for back-and-forth conversation, such as on a customer phone line. ADK usually stands for a development kit, though the summary does not spell it out. Details and pricing are not stated.

The phrase to notice is 'any team': building these agents should not be limited to a few specialists. Together with the Builder MCP, it shows PolyAI trying to lower the effort needed to ship a conversational agent.

What this means for you: if you run customer service or a small business with a busy phone line, read the original explainer to see what is actually required.

Proactive versus reactive. A regular chatbot is a shop assistant behind a counter. It waits for your question. A proactive agent is more like a colleague who notices your inbox piling up and asks, 'Want me to handle that?'

In general, it works like this. The agent keeps running. It watches only what you allow. It offers suggestions, and you approve or refuse. The word to keep is agent: an AI that acts, not just answers. The more it can act, the more the rules you give it matter.

Give your assistant a job description first. Works with ChatGPT, Copilot, Claude or any assistant.

1. Pick one small, repeating task.
2. Write four lines: the job, what you give it, what it returns, what it must never do.
3. Add: 'If unsure, ask me first.'
4. Run it a week, check every result, then widen the job.

Your one job: turn my weekly notes into five bullets. Never send or delete anything. If unsure, ask me first.

1. Open Chrome. Click the three dots at the top right, then Settings.
2. Click 'About Chrome' and let it update if it asks.
3. Back in Settings, find 'Safety check' and click 'Check now'.
4. On a sticky note, write: 'No website gets to make me paste commands.'
5. Tell one colleague or family member about ClickFix.

You will know it worked when Chrome says it is up to date and Safety check finishes. Updating is the best step you can take today.

Let a browser assistant read long pages for you. Check each company's pricing page for current costs.

This is an estimate. If you read four long pages a week and save about five minutes each, that is 20 minutes a week. Over a year, that is roughly 17 hours, about two working days.

What you give up: summaries can skip details. Read any part with numbers or deadlines yourself.

The mistake: putting off a security update because your site still works. Today's story shows why that goes wrong. Attempts to exploit the PHP flaw surged right after it was announced.

The fix: turn on automatic updates, or ask your host to confirm in writing that security fixes are applied within the week.

Cloudflare (a service that filters website traffic). Today's flaw story quotes its chief executive on surging attacks. Good for blocking junk traffic and staying up under load; it won't fix flaws in your own software, so keep updating.

Safe to use? Check their privacy page.

Worth it if... you run a small shop, club or blog.

What MCP is. PolyAI launched a Builder MCP. MCP is a shared standard that lets AI tools connect to other software, like USB-C: one plug shape, many devices.

Without a standard, every connection needed custom work. A company builds one connector, and any assistant that speaks the standard can use it.

The catch: each connection gives the AI more access. Connect only what the job needs.

An early MIT program called ELIZA was a simple chatbot that rephrased what people typed as questions. He was surprised by how many users felt it understood them.

I asked my always-on AI agent to take a day off. It added that to my to-do list.

• Tell a chatbot that waits apart from a proactive agent that speaks first.
• Spot a ClickFix trick and refuse to paste commands from a web page.
• Write a one-task job description for any assistant before you trust it.

Thanks for spending a few minutes with us. Come back tomorrow for more.

If this made you a little smarter, share it with one person who would find it useful.