Skip to content
← Public packages

@kentcdodds/closing-the-agentic-loop

Closing the Agentic Loop: Kent C. Dodds's talk and workshop deck. Default short talk; ?event=aie-2026 is the AIE hour cut with a Podium QR; ?format=full adds labs.

FollowForkOpen app

Files

03c-friction-checks.html

15 lines · 1.5 KB · HTML
<section id="friction-checks" data-title="Friction becomes static checks">
<h2>Friction becomes static checks</h2>
<ul class="big">
	<li>As static as possible: types, lint, tests</li>
	<li>Agents add a rule when something cost them time</li>
	<li>Only when you can't, bring in an LLM</li>
</ul>
<aside class="notes">
<p>Rung 4 and rung 7 territory: self-verification, and "if reviewers say it twice, encode it".</p>
<p>At BIOS SPHERE I opened Kody's validate script live (npm run validate, about 149 lines): type checking, end-to-end tests, linting, formatting, "all the static things that it can possibly do." The oxlint config has a whole set of rules agents added over time "because they're like, I wasted 20 minutes figuring this out and we can statically check that." Examples on screen: enforce import boundaries, no example identifier. "This is the first time I've ever looked at this file."</p>
<p>"You just turn these problems that the agents had into something static, whether that be in the types or in linting or in the testing or whatever. Like you want to have as much possible in the static checks as you possibly can."</p>
<p>The funnel (my answer to "how much LLM versus static analysis for PR review?"): "as static as possible with validation that runs automatically... And then only when you can't do that do you bring in an LLM." Even then, look for cheaper models.</p>
<p>What static checks can't cover (the app's actual behavior) is the control CLI's job, in pillar 2.</p>
</aside>
</section>