Context All the Way Down
Entire, and a simple proposition
A week ago (ish), a startup named Entire launched with a fairly simple proposition: store agent context tidily as part of each commit. I’m on board with that idea. In fact, I said in a previous post we should be doing this. There but for the grace of god, vision, speed and having previously been the CEO of GitHub goes me…
The meta-point: context isn’t a sidecar. It’s part of the artifact.
Anyway, right now I’m not-so-low-key obsessed with trying to nail down how software engineering (SWE) and hardware engineering overlap and differ as practices. Software is blazing the trail when it comes to AI impact, and so every other discipline should be asking itself how it measures up.
Six things SWE benefits from
I think about six things that SWE benefits from:
- Abundant open source code
- A historical record of the evolving intent that turned requirements into specs and then code
- Explicit interface contracts (e.g. typed APIs)
- First class version control, branching and collaboration flows (git)
- Continuous integration testing
- Cheap, quick sandboxes (e2e preview branches FTW)
Training data vs agent playground
You could roughly group these like so: 1 and 2 are training data; 3-6 make up the the ideal agent playground. For our current crop of model architectures, you don’t get off the ground without 1+2.
Hardware has some analogues here, but they’re fragmented across tools and vendors—and the seams are where progress goes to die.
Entire is basically our software friends jogging out ahead saying, “cool, looks like it’s gonna be context all the way down”. That is, point 2 continues to be really frickin important. Meanwhile in mech eng land we can’t quite remember how we built that last fleet of nuclear power stations…
The painful fact is, we built them, but we didn’t leave ourselves a replayable record.
The breadcrumb trail (and what hardware lacks)
There is something delightful about the way context is captured in software development. Code diffs tell a story and there’s an earnest dedication to high quality commentary (commit messages). At some point, SWEs agreed that leaving this breadcrumb trail was, on balance, worth the effort. Even though, when all goes to plan, it’s never looked at again.
On my mind right now—how do we nudge hardware engineering towards similar patterns of granular context capture? The next generation of tools, from CAD to system design, needs to solve for this.
If you buy this argument, the next-gen CAD/system tools need to treat provenance as a first-class output.
(Everything as code? 👀)