AI Agents Work Best When They're Not Fully

The counterintuitive thing about AI agents

There's a document making the rounds in the developer world right now that's changing how a lot of people think about building AI products. It's called 12-Factor Agents, written by Dex Horthy, and its central idea is surprisingly humble for a field that loves to overpromise.

The finding: the AI products that actually work in the real world — reliably, day after day — aren't fully autonomous robots making decisions on their own. They're mostly ordinary software, doing ordinary things, with AI stepping in at a few very specific moments where language and judgment genuinely matter.

Think of it like a good kitchen. Most of what happens there is process — prep schedules, mise en place, timers. The chef's creative judgment shows up at specific moments, not constantly. That's what good AI product design looks like too.

This repo also coined the term context engineering — now the dominant way developers talk about how you shape what an AI knows before it speaks. Instead of just asking AI a question, you carefully prepare what information it sees. It matters more than most people realise.

For founders, this is quietly reassuring. You don't need an AI that does everything. You need one that does the right things, at the right time, reliably.

Words worth knowing

AI agent — an AI that doesn't just answer questions, but takes a sequence of actions to complete a task on your behalf.

Deterministic — when software always does the same thing given the same input. Predictable, auditable, trustworthy.

Context engineering — the practice of carefully choosing what information you give an AI before asking it something, so it responds more accurately.

Production — developer shorthand for "working in the real world, not just in a test."

If you're evaluating AI tools for your business, ask vendors not just what their AI can do — but where exactly the AI is making decisions, and where it's following a fixed process.