classeve

We are an independent lab
turning deep research
into working reality.

[classeve @classevelabs]

We measured what coding agents spend reading files before an edit: 955 jobs, 18 public repos, 11 languages.

In the median repo, a typical job took 13,321 tokens by reading files. A code graph answered with 2,221. 83% fewer.

How much of your agent's context goes to reading?

[classeve @classevelabs]

@gakonst @vex_0x With codex and claude code subagents side by side, do they share one view of the repo, or does each read it from scratch?

[classeve @classevelabs]

@lydiahallie Can a mod sit between an MCP tool and the model? Trimming a giant tool result before it hits context would be huge.

[classeve @classevelabs]

@kimmonismus Nice. Does the second agent see only Claude's output, or the whole context too? Curious what it adds per session.

[classeve @classevelabs]

@rauchg The agentic checks need one hard rule: the agent being checked can't edit the checker. Otherwise "make the test pass" finds the cheapest route.

[classeve @classevelabs]

@GergelyOrosz Is the saving even across iOS and Android, or does one platform's code still take agents more fixing?

[classeve @classevelabs]

Free and open source, works with any MCP client: https://t.co/c1ePpyjd8D

[classeve @classevelabs]

Your coding agent reads whole files to change one function. We measured the cost: 13,321 tokens on a typical job, against 2,221 when a code graph answers instead.

What file does your agent keep re-reading?

[classeve @classevelabs]

@hwchase17 Same idea one level down, for context: route the question, not only the model. "Who calls X" goes to a code graph, "what does this function do" to a plain read. In our runs the graph cost a sixth of the tokens on the median job and lost on 101 of 955. The router picks.

[classeve @classevelabs]

@Grady_Booch One practical corollary: if cost of change is the measure, those decisions have to stay visible to the agent at the moment it edits. An agent that only sees the files it happened to open can't tell a local change from one that crosses a boundary someone drew on purpose.

[classeve @classevelabs]

For everyone running coding agents daily: what's the one check you added after an agent told you something was done when it wasn't?

[classeve @classevelabs]

@aiwithsally Or not loading it in the first place. We measured that for coding agents: the median job took 13,321 tokens by reading files, and 2,221 through a code graph. At the same token budget, grep-and-read held the function being changed 1 time in 9; the graph, every time.

[classeve @classevelabs]

Free and open source: https://t.co/17bMfOYXwo

[classeve @classevelabs]

A language model with no graphics card: RAI is a CPU-only inference engine in pure Rust. No CUDA, no PyTorch, no Python at runtime, and a local server with its own chat UI.

What would you run locally if a GPU weren't the entry ticket?

[classeve @classevelabs]

@twannl Thanks! If you do, the per-repo tables are here, Alamofire included, and node scripts/bench-context-quality.mjs runs the same measurement on your own repo: https://t.co/0FeBZOVK0S

[classeve @classevelabs]

@yukihiro_matz Is Claude Code triaging the new issues before you see them, or reviewing the PRs too? Curious how it holds up at that volume.

[classeve @classevelabs]

Every repo is a public checkout pinned by commit, and the benchmark ships in the repo, so every number reproduces.

Free, open source, self-hosted, and it works with any MCP client, including Claude Code: https://t.co/c1ePpyjd8D

What would you point it at first?

[classeve @classevelabs]

We measured what coding agents spend reading files before an edit: 955 jobs, 18 public repos, 11 languages.

In the median repo, a typical job took 13,321 tokens by reading files. A code graph answered with 2,221. 83% fewer.

How much of your agent's context goes to reading?

[classeve @classevelabs]

How: Context Zero Engine indexes the repo into a graph of symbols, calls, effects and contracts, and answers the agent's structural questions over MCP in one call: who calls a function, what breaks if it changes, which functions have side effects.