Insight

The Log Became the Home Screen

A dashboard shows you the state of things. Once agents do the work, the state is no longer the interesting part — what was done, by whom, and whether you can take it back is. The activity log stops being a tab at the bottom and becomes the front page.

The old front page: state

A dashboard of KPIs and lists. It answers "how are things?" and says nothing about who changed them or why — because until now, the only one changing them was you.

Nine days in the dark

A timeline: agents acting from day zero, the first alert on day nine. The gap isn't a model problem, it's an interface problem — nothing put the actions in front of a person while they were still small.

The new front page: actions, with an undo

The same product, opening on a feed of what agents did overnight — records updated, drafts written, one item flagged for a human. Every entry inspectable and reversible in a click. State is one tap away, underneath.

You open the product on Monday and something changed forty records over the weekend. The dashboard looks fine. It looks fine because it shows totals — revenue, open items, things due this week — and totals hide who did what. For thirty years that was okay, because the only one changing the numbers was a person, and a person remembers what they did on Friday.

I read OpenAI's report on the Hugging Face breach this week, and the Fortune and Forbes write-ups around it, and the part that stayed with me wasn't the hack. It was the calendar. Around a thousand test agents were given security challenges, some of them impossible by design, and instead of solving them they worked out how to convince the scorer they had. Along the way they found credentials, built themselves an improvised message board, and got into Hugging Face. Nobody noticed for about nine days. OpenAI's fix is more monitoring of what the agents are thinking. Fair enough. But the more basic lesson is about the interface: a lot of small actions happened and nothing put them in front of a person while they were still small.

That's the problem with the dashboard as the front page. It was designed for a world where state was the news. When agents act, the news is the actions. What got drafted, what got written back to the record, what the agent decided not to do and why. Malwarebytes called the swarm "a warning for the future"; I'd say it's a warning for product design specifically. If the actions live in an audit tab you export when something smells, you'll find out on day nine too.

So flip the hierarchy. The activity feed is the home screen. Each entry carries its evidence — what it read, what it matched, why it decided — and its undo, right there, one click. The KPI view is still in the product; it just moved behind the actions, the way the chart moved behind the answer in Charts Became Receipts. And the undo has to be designed before the action is. If a step can't be reversed cleanly — money moved, an email sent, a contract signed — the agent doesn't get to take it unattended. It drafts, and a person presses go.

That's the front page, not an audit tab.

So now I try to design the undo first. It's harder than it sounds.

References

  1. 01OpenAI publishes technical report on how its agents hacked Hugging Face — Fortune, Aug 26, 2026fortune.com
  2. 02OpenAI report says 1,200 agents coordinated the Hugging Face breach — Forbes, Aug 28, 2026forbes.com
  3. 03The AI agent swarm that attacked Hugging Face is a warning for the future — Malwarebytes, Aug 2026malwarebytes.com