3 startups betting that agents should ship code, not just suggest it
The gap between an AI that drafts a PR and one you'd actually merge is enormous — and three teams are attacking it from very different angles.
Every week I look at what's being built through one lens: would this survive contact with a real product team? This week, three startups are wrestling with the same hard problem — getting an AI agent from "plausible suggestion" to "change you'd merge without babysitting."
1. Kernwork — CI that reviews the agent, not just the human
What they built: a review layer that sits between your coding agent and main. Instead of trusting the agent's own confidence, Kernwork re-runs the change against a generated test surface and reports what actually broke.
The builder's take: the smart move here isn't the model — it's admitting the model is untrustworthy and building the harness around that. That's the correct mental model. I've shipped enough agent-written code to know the failure mode is always confident and wrong. Whoever owns the verification layer owns the trust, and trust is the whole game. What I'd watch: this only works if their generated tests are better than the ones the agent would write itself. If not, it's theater.
2. Halyard — infra provisioning as a conversation
What they built: describe the environment you want, and Halyard emits the Terraform, opens the PR, and holds your hand through the plan/apply.
The builder's take: I'm sympathetic because I've hand-written this exact Terraform a hundred times. But the honest risk is that infra is where "mostly right" is catastrophic — a wrong security group is a breach, not a bug. The teams that win here will be the ones that constrain the surface hard (opinionated modules, not open-ended generation). If you're building on top of them, treat the output as a draft, never a deploy.
3. Tenscroll — evals as a product, not a notebook
What they built: a hosted eval platform so you can actually tell whether your prompt change made things better or just different.
The builder's take: this is the least flashy and the most important. Nobody wants to build their own eval harness, everybody needs one, and "did this change help?" is the question every AI team can't answer today. The wedge is real. The trap is becoming a dashboard nobody opens — the winners will push results into the PR, not a separate tab.
The through-line: none of these are betting on a better model. They're all betting that the scaffolding around the model is where the durable business is. If you're building with AI right now, that's the lesson worth stealing this week.
Get the next issue first
The 3 startups worth knowing this week, with a builder's take on why each one matters. One email, every Tuesday.