Aigile: a method for building software with humans and AI agents

Software turns out right when misunderstandings are found early, and short feedback loops are the most reliable way to find them. Coding agents make each loop much cheaper to run, so it becomes practical to run more of them and to keep the written specifications small. Current spec-driven tools tend to go the other way, with large specifications and few feedback points, and in practice this works poorly. The assessment page explains why, with an example.

In every aigile setup, humans decide what to build, why, and whether the result is right. Agents can gradually take over the checking work, one area at a time, when a recorded track record supports it, and the work returns to humans when something goes wrong. Because the human part stays the same across all setups, a small team pairing with one agent and a largely automated delivery pipeline can follow the same method.

The method is written down in a working paper and nine deep-dive papers. Together they define thirty-six numbered rules, the protocols behind them, one worked example that runs through the whole series, known failure modes, and metrics. All documents are versioned and free to download.

Three tutorials each take you from installation to a first completed feature, on the same small example: one for SpecKit with its gated workflow, one for Claude Code with its always-on skill, and one that composes the method with an OpenSpec setup. Expect one to two hours for a full first pass, most of it real work on the feature.