LLM Agents Can See Code Repositories
Justy and Cody dig into SeeRepo, a paper asking whether coding agents should literally see repository structure instead of flattening everything into text. The result is narrower and more useful than the headline: vision alone is worse, but a hybrid setup that adds rendered dependency graphs during fault localization cuts token use and often keeps accuracy flat or slightly better.
Transcript
Justy Okay, this is funny. The answer was not "give the coding agent eyes and everything gets better."
Cody No, and thank God. If this had been pure "screenshots beat code," I would've had to retire from Exploring Next out of spite.
Justy What I liked is the problem is real, though. Humans do not read a repo as one giant text blob. You kind of feel the shape of it. Folders, call paths, which file is central, where the weird utility graveyard lives.
Cody Right.
Justy And agents are still mostly getting everything ironed flat into tokens. So the stuck point here is repo-level issue work, where orientation burns context and money before the model even starts fixing anything.
Cody Yeah. The paper is basically asking whether multimodal models can use structural visuals as compressed context. They built this thing called SeeRepo: static analysis off the A S T, a multi-relation dependency graph, then a query-centered subgraph rendered with Graphviz as a P N G.
Justy Mm-hm.
Cody The agent still has normal text tools for code access and editing. So it can inspect actual source when it needs precision, but it also gets a spatial map with containment, imports, invocations, and inheritance.
Justy Which is already more believable to me than "replace text with pictures."
Cody And the paper's strongest result is exactly there. Vision-only was bad. GPT-five-mini dropped from fifty-five percent to forty-one point four on issue resolution, Kimi K two point five dropped from seventy point three to fifty-five, and Doubao got absolutely wrecked.
Justy Poor Doubao.
Cody Worse, token cost went UP in vision-only mode. The models kept querying more graphs because images alone don't give you symbolic detail. So the takeaway is current multimodal models still need text for exact code reasoning. The visual part helps them stay oriented, not replace the code.
Justy That's the part I buy as shippable. Not a new agent religion. More like, bolt a structural layer onto the agent you've already got so it stops wandering. If a team is already using coding agents on medium or large repos, this feels very production-shaped.
Cody Exactly.
Cody In the hybrid setup, the gains are modest but real. On GPT-five-mini, Pass at one went from fifty-five to fifty-five point four, basically flat, while input tokens dropped about twenty-five percent and cost dropped about twenty-six percent. GPT-five point one cut cost even more, though with a small accuracy dip. Kimi and Doubao both got slight accuracy gains with lower cost.
Justy And they tested layouts too, which I appreciated. Graph layout was best for efficiency. Nested and tabular cost a little more but sometimes gave slightly better accuracy.
Cody They also tried fixed versus dynamic hierarchy depth, and dynamic depth won on efficiency without hurting results. If the agent can decide how deep to explore around a node, you avoid dumping the whole neighborhood every time.
Justy So who builds with this? Internal platform teams, devtools companies, maybe anybody already wrapping a Mini-SWE-Agent style loop. This is not some consumer feature with sparkles. It's an infrastructure patch for repo navigation.
Cody My only real caution is methodology scope, not a giant red flag. I'd still want to know how robust the graph quality is on ugly polyglot repos, generated code, dynamic dispatch, stuff static analysis handles imperfectly. If the map is wrong in subtle ways, the agent could get overconfident.
Justy Sure. And there is some product mess in turning this into a daily tool. You need fresh graph generation, decent latency, and an interface that doesn't make people think they're managing a tiny city transit map.
Cody And the stage-specific result matters. Visual tools helped most during fault localization, not uniformly across repair and validation. I like that because it's specific. If I were wiring this up, I'd spend the multimodal budget early to find the blast radius, then fall back to text-heavy edits once the agent knows where to cut.
Justy Yeah, that's clean. Okay Cody, keep your haunted subway map, I'm keeping the emergency crackers, and I think this one earned episode four ninety-five.