Rafael Mendiola · Co-founder and product engineer, okthink.ai

Loop Engineering by Rafael Mendiola

Loop Engineering is a code-first presentation about building reliable software with autonomous agents. It follows a real migration of a complex production application from Vite to Expo while preserving behavior, appearance, test identifiers, and the product contract. The central argument is that repeatedly prompting an agent supplies momentum, but it does not supply judgment. Reliable loops need an environment that helps the agent perceive the product, remember prior work, evaluate evidence, and stop only when the result is trustworthy.

What the presentation covers

The talk starts with Ralph loops, OpenClaw, and the skepticism surrounding long-running coding agents. It then shows why one-shot migration attempts and increasingly elaborate plans produced false finish lines. The successful process gave the agent a read-only reference product, an implementation target, a durable work ledger, executable parity checks, and explicit completion rules. As the agent takes on more implementation work, the engineer moves up a level: designing guardrails, evaluators, evidence systems, and the conditions in which autonomous work can be trusted.

Using this site

The public deck supports keyboard navigation, direct slide links, fullscreen presenting, and a print-oriented export view. Authoring tools such as annotations, speaker notes, slide inspection, and theme editing are intentionally limited to local development. JavaScript enables the presentation controls, but the summary on this page is intentionally available in the initial HTML for crawlers and clients that do not execute scripts. Agents can request this URL with an Accept header for text/markdown, read the curated agent index, inspect the sitemap, or discover the machine-readable HTTP contract.