Tooling is Rocket Fuel for AI Agents

What you’re seeing in SWE right now

If you are watching AI as applied to software engineering, you’re seeing the fruits of 2+ decades of developer tool refinement come together. The self-built innovations that have helped humans collaborate in the fastest, most explorative engineering discipline of this era are proving to be rocket fuel for the adoption AI agents.

Agents can do their best magic when the substrate is already magical: diffs, branches, tests, sandboxes, reproducibility.

The mundane stuff that matters

I’m talking the now-fairly mundane: transparent sharing mechanisms (APIs), version control, branches, end to end testing, ephemeral sandbox environments, and the fact these can all be neatly integrated and orchestrated. Software engineers wouldn’t tolerate anything less.

The lens for everyone else

So then a lens to view the rest of the engineering world through is: how the heck we gonna do this, folk?

My gut instinct: non-SWE disciplines won’t get much agent leverage by bolting chat onto legacy tools. Rebuilding the substrate with versioning, contracts, tests, and sandboxes will have to come first.