HarnessĀ²
The title does a lot of work here. Two steps, same verb, and the second one is where the money hides. Plenty of teams can ship a capable agent, but capturing the self-improvement loop, the part where every run makes the next one better, that's a whole different engineering problem. I think the teams that win will be the ones who treat the measurement layer as the product, not the agent itself.
The connection I keep coming back to is how early analytics startups ate the market by making the invisible visible. Same play here. If you can show a founder exactly where their agent improved and why, you've built trust that a demo never will. I'd want to know how they handle verification, because an agent that learns from its own mistakes can also learn the wrong lessons. That guardrail is the real asset.
I land on this being a genuinely useful framing for anyone building on top of LLMs. The agent is the entry point, and the self-improvement harness is the part that takes real engineering and real trust to get right.