26 posts
Human in the Loop
Notes on AI agents, building, and what still needs a human.
Latest
- Building with Agents The Feed Got Fast. Then It Got Noisy. Splitting our AI feed into two tiers fixed latency in February, but by October the two paths had turned into two competing editorial standards. I built a 50-item benchmark, tested a probability-based judge, and collapsed the pipeline back into a single builder.
- Human in the Loop What a Merged PR No Longer Proves A merged PR used to be rough evidence that someone understood the problem. With agents writing the code, that evidence is gone, and cheap code turns out to be the best tool for building it back on purpose.
- Human in the Loop The Agent Optimizes for What You Check People often complain that coding agents produce illegible code. Yet a repository merging 600 agent pull requests a day stays maintainable, and the difference comes down to the harness rather than the model.
- Human in the Loop The Interrupt Is the Work Models have improved dramatically, yet the quality of what a coding agent delivers still hinges on whether you intervene mid-run. This isn't just subjective impression: over-scoping and architectural bloat are real, measurable failure modes.
- Book Log Book Log: Escape from Freedom: A Theology Book That Read Like a Diagram of My Own Head Fromm's Escape from Freedom read like it had been following me around. It traces where the modern self came from, why work feels like proof of worth, and why the anxiety of choice is normal rather than a personal defect. It gave me real relief, and I recommend it to anyone wrestling with similar feelings.
- Building with Agents The Most General Tool Ever Built Sells One Specific Thing LLMs feel like they can solve every problem. But looking closely at ChatGPT, Claude Code, and Codex, product-market fit didn't come from generality. It came from aiming a general model at one specific problem customers eagerly pay for.
- Building with Agents The Agent Stack Is Standardizing in Real Time Standards usually arrive late, after technologies settle and winners emerge. In the agent space the order is inverted: protocols are being drafted in real time, in public, by fierce competitors. This post explores why that is happening and maps the emerging stack.