#5 🤖 When is an AI Agent the right call?

AI News

Top posts from this week

🛠️ An AI agent isn't a smarter chatbot; it's a loop that catches its own mistakes.

A chatbot answers once, but an agent works a task in steps: it acts, checks its own result, and retries when something's off, without approval at every move. On a real project pulling event listings from hundreds of partner websites, the first automated guess was wrong a third to nearly half the time, and it was that loop, not a smarter model, that made the result usable. Before paying for an "agent" feature, read Mateus's breakdown of when that complexity is worth it, and check whether your task is varied, frequent, and forgiving enough to need one.

🚪 Your AI provider can cut you off for reasons that have nothing to do with you.

OpenAI cut Cursor off after SpaceX bought it, and providers also drop whole regions and quietly retire models, on as little as five days' notice. Amanda walks through the four-column inventory OmbuLabs ran on itself, which company's model sits under each tool and what breaks if it says no, and which gaps are worth paying to close.

🔒 Handing an AI agent a password? There is now a safer way to do it.

When ChatGPT Work, OpenAI's workplace assistant, reaches a login page while doing a task for you, it now pauses and shows a separate form where you type your username, password, and two-factor code, so they never go into the chat. Ben's Bites walks through the new sign-in flow, including that the details go straight to the browser doing the work, that the session can stay signed in for later tasks or be cleared in settings, and that Ben still wouldn't start with your bank account.

📅 "Is it safe?" is the one question stopping most small business owners from using AI.

Join Fiona for a free 45-minute webinar on rolling out Claude in your business the right way, no technical background required. You'll walk away knowing what Claude actually does with your data, how to set ground rules your team will actually follow, and the first three steps to roll it out this month. Wednesday, September 9th, 11 am ET.

Stay tuned for more hands-on workshops too!

🐛 Can an AI model track down a bug your own engineers couldn't explain for years?

Anthropic says its new Claude Fable 5.1 traced the root cause of a crash in investment firm Millennium's internal systems that had stumped Millennium's engineers, and every other model, for several years. The release also cuts pricing for highly agentic workloads by up to 45% and adds Claude Mythos 5.1, a more permissively safeguarded variant reserved for vetted cybersecurity and life sciences work.

📊 A survey of 520 small firms shows what "cautious AI adoption" actually looks like.

​The Global Technology Industry Association's survey of 520 SMB decision-makers found 84% report a positive business impact from the AI tools they use, with half already using AI across operational and some customer-facing work while the rest deploy it tactically or run pilots first. Governance is lagging the usage, though: only 44% have a written acceptable-use policy, and 24% name data security, privacy, and compliance as a top barrier. Worth a look as a gauge of whether your own AI rules have kept pace with what your staff is already doing.

📄 Claude's memory now works across chat and Cowork, and you can edit what it keeps.

Starting August 25, Claude, Anthropic's AI assistant, uses a single shared memory across chat and its cloud-task tool, Cowork, so ask Cowork to draft an update for your manager, and it already knows who she is and how she likes them written. You can read, edit, or delete each saved topic under Settings > Memory, and sensitive subjects like health or politics stay off unless you opt in. For Team and Enterprise plans, an admin has to enable it before anyone can use it, so that is the setting to check if you want it.

Do you know anyone who would love to receive this newsletter? Tell them to subscribe here 👇 and catch the latest issue right in their inbox.