Interested in lots of things, but mainly crypto, browsertech, and interfaces.

It's Not an Agent if You Have to Ask (and that's okay)

April 17, 2025 ai agents

It’s Not an Agent if You Have to Ask (and that’s okay)

CoBot is marketed as “the first suite of AI agents built to accelerate your team.” Details are sparse, but their teaser video shows an app where you can install “cobots” for different tasks and then chat with them to get things done. In the example, a user asks the Stripe cobot to issue a refund, then asks the email cobot to draft a follow-up message. My first thought: is this just an MPC wrapper?

Maybe. But to be fair, there’s some real value in reducing the friction. Instead of switching between tools, you can just summon the cobot in context. That counts for something.

Still, this highlights a recurring issue with many agentic products: they’re not actually very agent-like. They’re reactive — sitting idle until you give them something to do — rather than proactive, solving problems before you even notice them.

There’s a big difference:


❶ Reactive

A customer emails me angrily at 3am. I wake up, ask the Stripe cobot to issue the refund, then ask the email cobot to draft a reply. Helpful? Sure. But if I hadn’t done anything, nothing would have happened. The bots are just waiting for orders.

❷ Proactive

A customer emails me angrily at 3am. While I sleep, CoBot detects the complaint, issues a refund, drafts a reply, and adds the whole exchange to a review queue. I wake up, see it handled cleanly, and move on. That’s agentic. It’s taking initiative.


I don’t have a strong opinion on CoBot itself — it’s just another entry in a growing crowd of agent platforms that feel more reactive than proactive. But that doesn’t mean it’s a bad product. In fact, most successful “agents” today are reactive.

Take Copilot, Cursor, Windsurfer — they’re not out hunting for bugs or rewriting your Jira tickets overnight. They wait for you to ask for help. And yet, they’re immensely valuable. Cursor might not do anything ChatGPT can’t do, but the experience is much more powerful when you situate the model inside a familiar environment. You don’t need to copy-paste anything. You get PR-style diffs and merge conflict resolution built-in. That difference in experience matters more than raw functionality.

So maybe CoBot and other agentic products don’t need to be truly proactive to be useful. If they reduce friction, fit into existing workflows, and save time, that’s a win. Fully proactive agents are the dream, but I’m okay with agents situated into my environment in the meantime.