Let's Go Christo!
#ai#agents#strategy

The Five-Question Filter for Agent Launches

There's a new agent product announcement every other day. Nate B Jones gave me a way to triage them in 30 seconds. I'm using it.

I subscribe to too many AI newsletters. So do you. The signal-to-noise on agent launches is getting worse, not better, as more companies pile in.

Nate B Jones published a five-question filter recently for evaluating agent announcements. I’ve started running every new launch through it before deciding whether to even read the post past the first paragraph.

#The five questions

  1. Is this infrastructure or a feature? Infrastructure goes underneath everything you build. Features sit on top of one workflow. Most “agent” launches are features dressed up as infrastructure.
  2. What state does it own? If the answer is “none, it just calls APIs”, it’s a wrapper, not a system. Wrappers don’t compound.
  3. Does it compose with what I already have? The right answer is “yes, in three obvious ways.” If composition is awkward, the tool will lose to one that composes cleanly.
  4. What does it cost at my scale of usage, not the demo’s? Demos run a hundred queries. I run thousands. The cost curve matters more than the per-call price.
  5. Who does this replace, and what would they be doing instead? If the answer is “nothing, it’s purely additive”, be skeptical. Real tools displace something.

#How I’m using it

I ran the last six things in my “agent launch” tab through the filter:

  • Three failed question 1. They were features. I closed the tabs.
  • Two failed question 3. They didn’t compose with anything in my stack. I closed the tabs.
  • One passed all five. I read the post, tried the tool, kept it in rotation.

That’s a 5/6 close rate in maybe four minutes total. The closed tabs would have eaten thirty minutes each if I’d read them.

#The deeper lesson

The five-question filter is not about agents. It’s about the layering principle Nate keeps coming back to: most agent decisions are about composition, not selection.

You’re not picking one winner. You’re stacking three or four specialized systems that talk to each other. So the question isn’t “is this the best agent?” The question is “does this slot into the stack I already have?”

That changes which tools get attention. The flashy ones with the impressive demos lose. The boring ones with clean APIs and predictable behavior win.

Hope runs on OpenClaw, delegates her chat through the Claude CLI backend, routes cheap calls through OpenRouter (MiniMax, Kimi, MiMo), and falls back to local Ollama for triage. Four systems. Each does one thing well. None tries to be the whole stack.

That’s the kind of setup the five questions select for. And once I had the filter, I stopped getting distracted by everything else.

Let’s go!

Share this post