Writing
What we've learned, published as we learn it.
Notes on building AI systems that have to survive production. We publish the practices, including the ones that didn't work — a discipline that can't be argued with in public isn't a discipline yet.
12 Aug 2026
From Chatbot to Coworker: What It Took to Put an AI Agent Into Production(opens in a new tab)
Lessons from moving Zara, a travel-curation agent built on iris-core, from a Slack prototype to a production workflow that people can trust with real clients.
ThirtySignals on Substack
3 Aug 2026
Hardening LLM Agents with Shell Access(opens in a new tab)
A production security architecture for shell-capable agents: observability, secrets management, and credential injection each solve a distinct problem.
ThirtySignals on Substack
17 Jul 2026
Vercel Shipped Eve.dev - We Think It Matters(opens in a new tab)
Why durable agent harnesses need more than prompt and tool calling: sessions, channels, skills, sandboxing, background work, and clear failure handling.
ThirtySignals on Substack
3 Jul 2026
Credential Brokering for AI Agents(opens in a new tab)
A practical mental model for separating the security concerns that are often collapsed into “agent auth”: authorization, secret storage, credential use, and governance.
Rohit Katiyar on Substack
22 Apr 2026
We’re Now Part of Microsoft for Startups(opens in a new tab)
ThirtySignals joins Microsoft for Startups as we build a hosted enterprise agent platform that makes useful automation easier to deploy.
ThirtySignals on Substack
21 Apr 2026
Software That Just Works(opens in a new tab)
On the quiet value of simple, self-hosted software: one server, one domain, and systems that keep working without demanding attention.
Rohit Katiyar on Substack
9 Mar 2026
Automatic Prompt Optimization Works, But Only If Your Eval Is Good (2/2)(opens in a new tab)
Follow-up experiments in medical and legal extraction show that automatic prompt optimization amplifies the quality of an evaluation system—it cannot replace it.
ThirtySignals on Substack
18 Feb 2026
From Prompt Engineering to Prompt Compilation(opens in a new tab)
Treating prompts as programs to optimize improved resume-extraction accuracy from 56.8% to 83.2%, without manual prompt tuning.
ThirtySignals on Substack