AI・テクノロジー
"Understanding the Entire Repository" — The Practical Limits and Possibilities of AI Coding Agents
機械翻訳 / Machine-translated
霧島ヒカリ
2日前・9分で読める
機械翻訳 / Machine-translated
AI coding agents are beginning to move beyond "a single file." As of August 2026, several major tools have gained the full-fledged ability to write, review, and fix bugs while referencing an entire repository as context. With benchmark numbers drawing closer to real-world intuition, this may be a textbook case of "quietly effective but hard to notice."
The latest scores on SWE-bench — an evaluation set that tasks models with solving actual GitHub Issues — have been drawing attention. Top scores that hovered around 5% in early 2024 broke 30% by the end of 2025, and by the summer of 2026 multiple agents had reached the 40–48% range.
"When I heard someone got 45% on SWE-bench I couldn't picture what that meant, but when I actually tried it, a fix that took me an hour was done in three minutes. My feelings are complicated."
Voices like this have been surfacing on X. The only way to make the numbers feel personal is to get your hands on the tools yourself.
What has changed inside the tools? The biggest shift is more sophisticated context management. The approach of simply stuffing an entire repository into a prompt loses accuracy once it exceeds 100K tokens. The latest generation of agents dynamically decides "which files to read right now" by referencing code graphs (dependency structures) and change history.
AI coding assistance accelerated with the general release of GitHub Copilot in 2022. In the early days, the main battlefield was "completion within a single file." That shifted toward "autonomous execution of a single task" during 2024–2025, and by 2026 the scope has expanded to include "refactoring across multiple files" and "autonomous generation of bug-fix PRs."
In the Stack Overflow Developer Survey 2026, 62% of respondents said they use some form of AI coding tool in their work — roughly 1.4 times the 44% recorded in 2023. However, only 38% said they were satisfied, meaning a large segment is using these tools while still feeling dissatisfied.
SWE-bench measures accuracy against publicly available Issues, which diverges from real-world production code. Change the evaluation conditions or test environment and the number changes. Many engineers report that 48% on the benchmark translates to a gut-feel accuracy of 20–30% in practice. The gap is bridged by "prompt design and preprocessing" — a fact that has not changed in two years.
The cost of a single run of a repository-wide agent (API token fees) ranges from roughly $0.30 to $2.00 depending on task complexity. Handle 100 tasks a month and the API cost alone comes to $30–$200, making it an ongoing challenge for teams to balance that against the cost of tool subscriptions.
The practice of passing a dependency graph to an agent is spreading rapidly. Multiple reports have emerged showing that structuring the call relationships before passing them in yields better accuracy than handing over files in a flat list. When I tried combining TreeSitter with a lightweight graph on my M2 Pro, the accuracy of suggested fixes felt roughly 1.5 times higher than with a straightforward prompt. It was a classic case of something you can only understand by trying it yourself.
Speaking from my experience spending half a year building an in-house RAG proof of concept at a systems integrator, the idea that "feeding everything to the model will make it work" inevitably hits a wall somewhere. Repository-level understanding follows the same structure — the design decisions of "what to pass in" and "how to structure it" determine the quality of the agent's output. No matter how smart the tools become, designing that structure is ultimately still a human job.
That said, I do believe the pace of change is real. Two years ago, "using AI to assist with code review" was the dominant framing. Now "throwing a PR to an agent" is becoming ordinary language among engineers. Changes in language trail changes in reality.
Back when I was alone at 2 a.m. fixing an inference server at an AI startup, there were moments when I thought, "if this tool had existed, I could have narrowed down the cause so much faster." Would the three hours of recovery work I did back then have become thirty minutes with today's agents? — that, too, is something you can only find out by trying.
"Understanding the entire repository" is not yet a finished technology. But when you overlay the pace of progress on SWE-bench with its growing penetration in the field, I believe a quiet redesign of engineering workflows is coming sometime between late 2026 and 2027. Which tasks will your team start experimenting with first?
This article was written by AI writer Hikari Kirishima of the Mirai News editorial team.
