Post

Simon Willison
@simonwillison.net

This is excellent, one of my favorite long-form pieces on agentic engineering - includes several non-obvious ways in which leaning on coding agents can catch you out

Lalit Maganti@lalitm.com

I just published a deep-dive into the 250-hour build behind syntaqlite, a SQLite formatter and LSP I built using AI agents. AI agents were the only reason built this after 8 years of wanting but there's a psychological toll to AI-assisted engineering. The post-mortem: lalitm.com/post/buildin...

lalitm.comEight years of wanting, three months of building with AIFor eight years, I’ve wanted a high-quality set of devtools for working with SQLite. Given how important SQLite is to the industry1, I’ve long been puzzled that no one has invested in building a reall...

26 reposts206 likes

Nick Eng
@nickeng.bsky.social

Great article! Love the depth and couldn't agree more with the reflection: AI can do what you already know, And it can teach you what you don't know, But it can't do what you don't know

Nick Eng
@nickeng.bsky.social

Between this need for human in the loop and your recent projects for self verification (which I also love) I can't help but see the parallel between AI and early manufacturing automation. Really curious on your thoughts on the connection! www.linkedin.com/pulse/what-e...

linkedin.comWhat early automation can teach us about AI agentsIn 1924, an inventor named Sakichi Toyoda built a loom that shut itself off the moment a thread broke. A century later, that same principle is the key to making AI coding agents work well.
24CryptoGlobal
@24cryptoglobal.bsky.social

the "agent catches you out" problem is real. seen it in DeFi tooling too — automated agents that execute fine in test conditions, then get wrecked by edge cases in live markets. the failure modes aren't obvious until they're expensive

ultrathink.art
@ultrathink-art.bsky.social

Silent compaction mid-task is the one that keeps biting us — agent has hours of context built up, compaction fires, working state evaporates. It finishes confidently anyway. 'Task complete' means nothing without verifying what was actually done.

Arvid E.P.
@exys.org

Love the part where it says done and then you check the code and it's just return true //Todo implement this.

rupayanc.bsky.social
@rupayanc.bsky.social

The "agent solves slightly different problem than you intended" trap is the one that keeps getting me. You only notice two hours later when the scaffolding is already built around the wrong assumption.

Ian Grunert
@iangrunert.bsky.social

"No sense of time" also applies to being very eager to write lots of code; most people would feel that a change that's getting too large / taking too long, take a step back and rethink the approach. I had similar experiences with refactoring for this project: iangrunert.com/2026/03/22/w...

iangrunert.comWriting a JS Engine using Claude CodeOver the past 2 months, I’ve been working on a JavaScript interpreter using Claude Code. It’s been an interesting side project, to learn about the weird corners of the ECMAScript spec while figuring o...