---
title: "Recursive AI Research"
url: https://recursiveai.org/research/
description: "Notes from building software with a fleet of agents. Patterns that survived contact with real work, most of them useful well beyond our own tools."
---

# Notes from building with agents.

Patterns that survived contact with real work. Take what’s useful; most of it applies well beyond our own tools.

- [RSS feed](https://recursiveai.org/research/feed.xml)

## All notes

- [Normalize like a skeptic](https://recursiveai.org/research/normalize-like-a-skeptic/) (On measurement): Measuring AI-assisted development honestly took more design than expected. Score scope, never lines; classify diffs, never messages; decompose, don’t zero; and drop any era with known-missing data from headline claims.
- [Pixel statistics cannot see meaning](https://recursiveai.org/research/pixel-statistics/) (On evaluation): A character passed three experiments and a cross-experiment comparison with no face. Mechanical metrics validate geometry, scale and consistency well, and validate semantics not at all.
- [Corrective feedback should be git-diffable](https://recursiveai.org/research/git-diffable-feedback/) (On teaching agents): When an agent misbehaves, the durable fix is a short, readable rule appended to its definition. The learned policy is text you can read, diff, blame and revert.
- [The best sub-agent output is a tool call, not a report](https://recursiveai.org/research/sub-agent-tool-calls/) (On context): A helper agent burns its own context on messy discovery, then injects clean, executed tool calls into the parent’s queue. The parent’s context holds the answer, not the search noise.
- [Decomposition delivers what checkpointers simulate](https://recursiveai.org/research/decomposition-vs-checkpointing/) (On durability): When every step is an independently addressable endpoint, replaying step 30 means calling step 30. And when a multi-agent build needs resuming, the resumer should be an agent, not a replay engine.
- [Separate orchestration from the unit of inference](https://recursiveai.org/research/orchestration-vs-inference/) (On orchestration): A flow is deliberately not a workflow engine. Control flow belongs one layer up, in real code. Fusing the two couples the thing you want to iterate on to the thing you want to be boring.
- [Catalog size measures the wrong thing](https://recursiveai.org/research/integration-catalogues/) (On integrations): A connector library is a moat only against users who can’t generate their own tools. The axis that actually matters is time to the first working integration.