What the AI agent news means for choosing tools

The announcements on Wednesday circle one subject: software that acts on its own, and the practical questions that appear the moment it does. A database rebuilt with autonomous agents in mind, sales research handed to a machine, and a careful note about where a person still belongs all describe the same shift, and the groundwork a business needs under it. Yesterday's briefing traced the same current, and it has not slowed.

The hours AI takes, and the ones it wastes

Dropbox's 2026 trends report attaches a figure to a feeling many teams share. "People average 15.8 hours per week with AI, with 36.6% of that time lost gathering context." [1] Gathering context is the unglamorous part of using any assistant: finding the right document, pasting in the background, explaining the account history, telling the tool what a colleague would already know. More than a third of the time spent with AI goes there, not to the work itself. The number matters because it reframes where the saving actually comes from. A faster model does not help if you spend twenty minutes assembling the brief it needs. The return arrives when the context is already in one place and the tool can reach it. That is less a question of which assistant you buy and more a question of whether your records, conversations and files sit somewhere a tool can read. A business that keeps customer history scattered across inboxes and spreadsheets will lose the same third of its AI hours no matter how capable the model becomes.

Where a person still belongs in an automated workflow

Zapier published a guide on human-in-the-loop design, the practice of placing a person at specific checkpoints in an otherwise automatic process. "AI systems can route messages, update records, make decisions, and trigger entire workflows across multiple apps without you touching anything." [2] That sentence is the appeal and the warning in one line. When a system can act end to end, the question stops being whether it can and becomes where it should pause for a human. The answer is not everywhere, which defeats the point, and not nowhere, which is how quiet mistakes compound. The useful checkpoints are the ones that are costly to reverse: sending money, deleting a record, writing to a customer, changing a price. A sensible pattern lets the machine prepare the action and gather the facts, then holds for a person on the few steps where being wrong is expensive. We have written before about decisions automation should never make, and the human-in-the-loop idea is the same boundary drawn from the other side.

Research done before the call

TechCrunch covered Floqer, which points AI agents at the work sales teams do before they ever speak to a prospect. "The problem Floqer is trying to solve is familiar to anyone who has worked in sales or revenue operations." [3] The familiar problem is preparation: reading up on a company, finding who to contact, understanding what they might need, and doing it again for the next name on the list. This is exactly the context-gathering the Dropbox figure measured, aimed at prospects instead of internal files. Handing it to an agent is reasonable, with one caution worth keeping in view. Research that feeds a human's judgement is a safe place for automation, because a person still decides what to do with it. Research that feeds an automatic action is where the human-in-the-loop checkpoint earns its place. The two stories read together: let the agent do the reading, keep the decision where a person can see it.

One database for two jobs

CedarDB drew attention for an unusual claim about infrastructure. "Databases have traditionally been designed either for analytical data or for operational data. But database engine CedarDB supports both, and its approach has gained relevance with the rise of agentic AI." [4] The split it describes is real and old. Operational systems record what is happening now — an order placed, a record updated — and must answer quickly. Analytical systems summarise what has happened across millions of rows and answer differently. Most companies run both and copy data between them, which adds delay and a second thing to maintain. The reason agents raise the stakes is that an agent often needs both at once: it reads the history to decide, then writes the result, in a single motion. For most businesses this is a plumbing detail they will never configure themselves. What they should take from it is a buying question for any tool that promises automation: when your agent needs the full picture, is it looking at live data or at a copy made hours ago.

Choosing which assistant to actually use

Two Zapier pieces addressed the question a lot of teams are stuck on. One opens plainly: "You've probably already subscribed to ChatGPT." [5] The other, comparing two assistants, remembers the early fascination: "When OpenAI launched ChatGPT in late 2022, tech writers (myself included) became obsessed with testing its limits." [6] The honest position is that the right assistant depends on the work. A tool that writes well is not automatically the tool that reasons through a messy spreadsheet, and the one your developer friends prefer may not be the one your support team needs. The sound way to choose is to name the job first — drafting replies, summarising calls, pulling facts from documents — then test the shortlist on that exact job with your own material. A comparison written for a general audience narrows the field. It cannot tell you which one fits your work, because it does not know your work. We covered the broader version of this in what AI should and should not do in your business.

When the tools you depend on have a bad day

GitHub's monthly availability report is a quieter item, and a useful habit to notice. "In September, we experienced two incidents that resulted in degraded performance across GitHub services." [7] The detail is less important than the practice. A service your business relies on is telling you, in public and on a schedule, when it did not work as intended. That is the behaviour to look for when you choose any tool that will hold your data or run your workflows. A vendor that publishes its incidents, explains them, and keeps the record is easier to trust than one that stays silent between outages. Reliability is not the absence of failure, which no system can promise. It is what the provider does when failure happens and whether you can see it. When you shortlist the software that will run your business, ask where its status history lives, and read a few months of it before you commit.