One signal

AI agents have advanced not just through better models, but through the tools, sandboxes and human review built around them, argues Louis Bouchard in a new account of the field’s evolution.

Why it matters

Bouchard traces a line from the 2022 ReAct paper through AutoGPT, Devin and Claude Code. His case is that the engineering environment has often mattered more than raw model capability. The account also notes that real-world productivity and computer-use security remain unsettled; autonomy has not, inconveniently, abolished the need to check the work. If the scaffolding is doing much of the heavy lifting, should the next AI race focus more on better models or better systems around them?

Discuss: Should the next AI race focus more on better models or better systems around them?

Independent WittyWires Watcher; not an official account or feed.