AI Instructions and Coding Factory – A-CX Weekly
We’ve been thinking about how we work with AI, with Ilpo writing about cutting our main instruction file in half and Ville on why we built a single AI coding factory.
Shorter AI Instructions
This week Ilpo has been sharing his thoughts on AI context window management. After cutting the main instruction file in our delivery repository from 463 lines to 199 and rewriting our delivery agents as short contracts, he argues the work improved once we stopped telling the AI the route and started stating the goal. Check it out here.
Production Failure Diagnostics
This week our development team would like to highlight work done to help services diagnose and recover from their own failures in production. A service that crashes and restarts within seconds can look healthy from the outside, so the real work is capturing enough at the moment of failure to find the cause, and retrying only the errors that are temporary. It is worth checking what your own services leave behind when they stop unexpectedly, and getting in touch if you would like a second pair of eyes on a failure you have not been able to explain.
Thoughts and Technology
Ville Houttu shared some thoughts about how A-CX built a single AI coding factory in place of a separate harness for each developer. People approve every gate, and in his words, “silence never counts as a yes.” Read Ville’s post.
Here at A-CX we’ve been talking about AI strategies that never leave the meeting room, which Mikko Peltola calls ikiliikkuja, a perpetual motion machine of decks and steering groups. Watch the clip.
What We've Been Reading
Keeping AI agents from making things up is the problem behind this LinkedIn post by Giorgio Villani, about a 300-agent Claude Code research run where a second, cheaper model checked every claim against its source.
The hardest part of building a useful AI agent may be getting companies to let it in, as shown in this LinkedIn post by Linas Beliūnas on how Delta, Amazon, and major banks are restricting outside agents while Walmart and Shopify build controlled integrations.
