One thing I keep hearing around AI coding tools is that engineering leaders want more velocity, while engineers are increasingly worried about what that velocity actually looks like.
Nobody wants to spend their day pushing slop.
A team can open more PRs than ever and still end up spending a surprising amount of time reviewing weird diffs or asking why the agent changed something that seemed fine before.
And if enough of those changes make it through, customers eventually notice. Bugs show up and small behaviors change. Things feel a little less reliable for everyone.
When that starts happening, the obvious reaction is to add more guardrails. Catch the bad diffs before they merge. Or pull back and let agents write less code.
There's another option we're more interested in: make the agent less likely to produce the bad diff in the first place.
The model can often write perfectly reasonable code. The problem is that it has no idea what happened six months ago at your company.
Say there's some ugly authentication code. An agent looks at it, sees an obvious cleanup, runs the tests, and opens a PR. Looks good.
Except six months earlier, an engineer spent an afternoon debugging an auth issue that only happened for one customer with a strange SAML setup. They tried the cleaner approach already and eventually landed on the ugly workaround.
None of that is obvious from the code.
Maybe pieces of the story are in a PR, Slack thread, or an old agent session. More likely, some of it was just typed into the agent's prompt while the engineer was debugging.
The next agent doesn't know any of that. This is something we're working on with OliverGraph.
We're trying to save what happened during the work, not just the final diff. If another agent touches that part of the codebase later, it should know what the previous engineer and agent already learned.
When an agent starts working on something, it shouldn't just retrieve relevant code. It should be able to retrieve some of the history behind that code too.
Something as simple as:
This workaround came from an incident three months ago. Someone already tried removing it and hit a race condition.
That's the direction we're betting on.
Teams are going to keep using agents to ship more code. I'd rather help make those diffs better than build increasingly large filters to catch the bad ones afterward.