November 10, 2026 · Agents
The agent demo works. The agent doesn't.
Nineteen agent tools in the library, and the ones still running six months later are the ones that only read. The gap between the video and the deployment is three gates, and all of them are boring.
Every agent demo I have watched this year works. That is the problem.
The pattern is always the same. A clean browser, a logged-in account, a task like "find the cheapest supplier for this part and draft an email". The agent clicks through four pages, reads a table, writes a paragraph. Ninety seconds. Applause.
Then you try it on your own machine and it dies at the login screen.
We now have nineteen agent-category tools in the library, more than we have in coding. That number tells you where the money is going, not where the value is landing. In the export businesses I visit, the agents that survived contact with reality are narrow, boring and read-only.
Three gates kill most of them
Credentials. Your supplier portal wants a username, a password, and a code sent to a phone that belongs to the boss. The agent can hold a password. It cannot hold a phone. Every workflow that touches a real system eventually hits an authentication step designed for a human with a pocket.
Error handling. A demo has no 500 errors, no "session expired", no PDF from 2009 that renders as an empty white box. A production agent needs to know when to stop, and stopping is the hardest thing to teach something trained to keep going.
Audit trail. When the agent books the wrong container or emails the wrong price, someone has to reconstruct what happened. If the reasoning lives in a chat window that scrolls away, you do not have a process. You have a story.
What is actually running
A browser agent that checks three supplier portals every morning and posts a screenshot to the team channel. Read-only. Nobody trusts it to order, and nobody needs to.
A crawler that pulls competitor catalogue pages into a sheet once a week, so the sales team stops asking "what are they charging now" in the group chat. That is Firecrawl, and it is the least impressive tool we list.
A coding agent that writes the throwaway scripts — rename these four hundred files, convert this CSV — that used to eat a Friday. OpenHands does this well because the blast radius is a folder on a laptop.
None of these are impressive in a demo. All of them save hours, because they fail loudly and cheaply.
The rule I would give anyone starting
Let the agent read, keep the human on the write. Every step that changes money, stock or a customer record stays manual until you have watched the agent fail at it fifty times and know exactly how it fails.
That is less exciting than the video. It is also the only version that is still running six months later, and the only one your insurer will recognise.
A note on independence. We do not take vendor money for coverage — the server is paid by readers, which is why a post like this can say the demo is theatre. If it saved you a bad pilot, you can chip in here; if not, the next one is free anyway.
We publish the pricing model, minimum commitment and export-readiness notes for every tool in the library, and we flag it when a vendor changes one without changing the headline price.