Anything "for agents" needs to provide some kind of evidence it's better than what the agents already have baked into the model training data. It can't just be "easier" on some dimension, because the model has already learned the hard parts of the old thing and models can't make new memories to learn new things, so there is always a context cost for the new thing.
Models know git because there's a monstrous amount of git in their training data. Models never heard of a new thing "for agents", so you have to teach them to use it via skills and docs. Models can, of course, follow documentation, so there's nothing stopping them from using the new thing...but, the new thing "for agents" starts the race well behind the known thing that was built for humans a decade or two ago and has huge amounts of training data baked into every model.
I'm not saying nobody should make new things (an accusation I've gotten when saying something similar about a previous "for agents" thing), of course people should make new things. I'm saying that when I see "for agents", I think, "prove it". Agents don't have trouble with git, so there's gotta be some kind of pain point about using git with agents that I'm unaware of that this solves somehow (but isn't expressed on the page) or this isn't actually for agents, it's just a project someone wanted to do (and that's also fine!). But, if the latter, "for agents" is merely marketing and I'm not interested.
I'm not sure I understand this argument. I create new tools all the time as part of my development work, and I have skills stored that tell agents how to use them. They use them flawlessly.
When I say "benchmark the query engine using the foobar dataset and compare it to run 431", the agents go and run my special benchmark tool and use the different subcommands to compare results and so on.
I'm sure a new VCS would be a little less smooth sailing, but not by much.
I think the issues is, it is going against a very well established pattern. I have a tool that wraps ripgrep so that search results always includes context and from time to time, the agent will use ripgrep by itself and when I ask why, it would go "yeah I should have done that"
There are work arounds though and I am creating what I call knowledge triggers for Pi that are similar to claude's "PreToolUse" so having the agent use oak all the time is not an issue in my opinion.
The challenge for oak is why? Considering how I actually want to slow agents down so I can ensure it is doing the right thing and because the massive bottle kneck is the LLM themselves, speed when measured in milliseconds or even seconds will not concern many.
I thought oak was more of, we know how to prompt inject context based on code that is stored in oak for example, but faster operations can help, but the use case is limited. The missing piece for better/correct code is context at the right time.
> Models know git because there's a monstrous amount of git in their training data. Models never heard of a new thing "for agents", so you have to teach them to use it via skills and docs.
Another option: when model invokes standard tool, rewrite the invocation to newfangled tool.
Bunch of ways of doing it:
(a) Invocation of standard tool returns error saying to use newfangled tool instead
(b) Invocation of standard tool returns message saying it has been dynamically rewritten to invoke newfangled tool, followed by newfangled tool output
(c) Invocation of standard tool in context is dynamically rewritten to invocation of newfangled tool, prior to execution
In case (c), the model ends up thinking it somehow knew about this new thing all along, even though it actually didn’t
Options (a) and (b) add more bloat to the model’s context window and option (c) seem to reduce to having similar functions that already existed. There is also the option to trick the LLM that it’s using the old function exactly as-is, while the harness abstracts away a completely different methodology. Cursor often does exactly this: they use an internally built vectorized search when the model calls the default “find” bash command. The LLM is none the wiser that the function’s implementation is completely different.
Regardless, in any of these cases, the implementation for any of these above options may be vastly superior to the “naive” implementation for agents — but then the parent comment here is right that an engineer would need to justify their implementation to users, not just make a loud conjecture. It’s a non-trivial claim to say that a bespoke solution not present in tool-use training and accounting for context-rot would result in a better performing model. Moreover, justifying an agent-specific efficiency gain that humans wouldn’t benefit from makes the claim even more non-trivial. Using Sagan’s razor, it’s then reasonable for people to ask for a comparably non-trivial amount of evidence.
Totally correct on the burden of proof here. Agents DO know git extremely well. There’s a huge amount of git in model training data, and anything new starts behind because you have to teach the model what it is, what commands to run, and where the sharp edges are. For us “for agents” does not mean “new syntax that we hope agents can read docs for.”
The thing we’re trying to optimize is not whether an agent can remember the command. It’s the runtime shape of agent-driven development.
When an agent drives a VCS through a captured terminal, things that are tolerable for humans become direct costs: clone/setup time, worktree setup, full status output, huge diffs, branch cleanup, interactive prompts, shared-checkout mutation, repeated preflight checks. Those costs show up as wall time, bytes over the wire, transcript tokens, and recovery steps.
So the Oak bet is narrower than “agents can’t use git.” They can. The bet is that if you assume branch-per-agent workflows, lots of parallel sandboxes, large repos, and non-interactive command execution, the VCS interface should have different defaults if you want to optimize for shipping speed and efficiency of token usage. If you're already going fast enough and not running out of tokens - then using oak seems pretty silly.
People do not need to ditch git to try Oak out. One workflow we care about is letting agents work in Oak where the agent-specific costs matter, then exporting back to git for the human review, CI, release, or compliance workflows.
Totally agree this should be provable and benchmarked. The homepage has Oak vs Git numbers because we do not want “for agents” to just be vibes. We’re measuring transcript bytes, estimated tokens, tool calls, wall time, large diff/status behavior, and contention in agent-style workflows. We’re also working on the benchmarks repo in the open: https://oak.space/oak/benchmarks
The exciting part to me is that we can already improve on tokens and timing despite starting with the model-prior deficit you’re describing. If we can win on measured agent workflows while git still has the advantage of being deeply baked into the models, I’m incredibly bullish on where Oak can get to as the tool and the ecosystem matures.
Longer term, if Oak proves useful and sticks around, future frontier models will likely have more Oak examples in training data, which lowers the upfront learning tax for an extra boost.
How did you speed up things (eg clone, worktree setup) compared to Git? Could the same work for human facing tools?
[dead]