First of all, the whole readme section about benchmarks appears to be Claude/Codex written. What a slog to read this.
Second, they claim success on SWE-Bench Verified, but it's only on 50 tasks, not making clear how these tasks are chosen. I know from experience that you can keep selecting sets of tasks until you get a result. Also, this was run 1x, and the lift they show actually only has a p val of 0.22.
There's a famous book "how to lie with statistics". I don't think the continual posts about benchmark results here are purposeful lies, but I think it's just so easy to fool yourself (and I've been burned, most recently building http://pellmell.ai ). Also I don't think this post is the most egregious example.
One concern I have is that right now each session gets fresh "eyes" on the problem. Right now I find I get a lot of mileage out of a combination of long-running sessions and fresh ones. I worry with a single generated concept graph that gets only incremental refreshes will become stale slowly, and in subtle ways that are hard to detect. That could lead to semantic drift in the graph from reality, and every new session will take the drifted form as gospel. Have you run any long tests (weeks or longer) on this to make sure that this doesn't happen? My understanding is SWE Bench is only a point-in-time evaluation.
Also the graph is stored in the repo, right? How mergeable is it? I know I wouldn't want to do conflict resolution on that myself, and even Opus struggles to keep all the references correct (especially when comments are involved) when there's an B->C, A->B symbol rename.
We have run tests on DeepSWE as well which are long running tasks, we got 20% better accuracy on the tasks where sonnet 5 failed. didn't want to post that numbers yet as I think we can do better on DeepSWE and on a cheaper model like gpt-5.6-luna or grok-4.6
Maybe this thing is great, but it cannot be determined with this presentation.
And may your deity of choice help you if you decide to venture on subjects which you don't have a deep understanding of, as half the sentences it generates will be (barely) cohesive.
It probably was, at least inadvertently. Very easy to imagine that one of the post-training step (RLHF, DPO etc) reinforced the "sounds clever" behaviour.
> Efficiency is a 162-run controlled benchmark (same agent, same file tools, only the context differs).
It's so bad that I can tell it's Claude, specifically Opus 4.8/5.0, from the first 6 words.
so graft solves for making claude code understand your codebase every session.
And I feel you can save a lot more tokens and also increase accuracy by just switching to new claude session without worrying about your codebase context. (doesn't solve for the session context yet, but we are still figuring out how can we solve for that. Maybe hooks that can manage your CLAUDE.md)
Also, for java projects we do support a more deeper graph, a few of our contributors are working on making it better for java projects, let us know if you have any ideas there.
You feel it while working too. graft hands the agent the exact file and line it needs for each question, so Claude keeps using graft.
Same is true for any other cli tools, claude never actually uses them as it's trained to use grep. but for graft as we set the directive to use graft at the start of the session and before the turn, claude just knows graft exists and also get the relevant context without it going and calling tools it was not trained on.
Hooks also solve for the issue of the LLMs not calling our CLI tools instead of Grep.
Every task, your coding agent starts blind. Before it changes anything, it re-explores the repo: grep a term, open a file, follow an import, back out, try again. It is rebuilding a picture of a codebase it mapped an hour ago and threw away.
That rediscovery burns most of a run's tool calls, tokens, and latency, and it is pure overhead"
The author of this article brings up a very interesting problem -- that, at least as far as using LLM's as coders/coding assistants go, eventually context runs out and context related to the underlying codebase does too. This in turn burns tokens and in turn, wastes energy resources.
Historically (well, in the past couple of years!), a bunch of solutions have been proposed to address this problem (i.e., take abstracts/subsets/maps of code, write them to different databases and persistent storage methods, bring them back in when the LLM requires it, etc., etc.)...
But there's no really good solution to this problem (although, arguably Graft goes a lot farther than past tools and should be commended for that!) because the problem seems to lie in separate parts, across several problem domains:
1) LLM context window size -- limited. Anything that future LLM's do to make context windows larger will help ameliorate this problem.
2) Lack of a good way to represent a codebase to an LLM for training other than text.
In other words, first we need some kind of way to map codebases into Tensors rather than text (i.e., a higher-level "map" of the code) then train future LLM's on those code-specific Tensors.
3) Arguably, programming languages themselves share some of the blame...
Programming languages have historically been written so that an arbitrary corpus of text represents and can be interpreted and/or compiled into a computer program.
That is, while tools for mapping codebases exist, tools for directly training LLM's on those specific created "code maps" as Tensors, do not, do not seem to, or at least I'm currently unaware of any!
(Anyway, just thinking aloud...)
Graft looks good, and looks like it has made some serious inroads to solving the problem...