Beyond grep: The case for a context rich AI coding harness

Asteria This is one of those arguments that sounds like plumbing until it decides whether a coding agent feels magical or just very expensive grep.

Draco Yeah. The real claim is not that semantic search is cool. It is that model intelligence and codebase context are separate inputs, and the harness decides how much time the model wastes finding its way around.

Asteria Right.

Draco Augment's bet is that a lean harness leaves too much exploration inside the token bill. Claude Code's team has made nearly the opposite bet: keep the harness light unless structured context produces measurable gains.

Asteria My week has been weirdly full of people treating every new agent wrapper like a constitutional amendment. So I appreciate an argument about the unglamorous part that might actually change the bill.

Draco Same. I have spent enough time staring at claims about agent autonomy that a vector database now feels almost restful.

Asteria Ha! That is bleakly soothing.

Asteria Augment pre-indexes a repository with embedding and retrieval models, then stores that representation in a vector database. When an agent gets a task, the system pulls conceptually related code instead of making the model wander through files with grep alone.

Draco Mm-hm.

Asteria They say the backend retrieves in sub-milliseconds. More importantly, Perneti argues this matters most in big private repositories, where the model has never seen the code and cannot lean on whatever it absorbed from public projects.

Draco That part holds up technically, with an asterisk. A retriever is another prediction layer: it can surface the right architectural neighborhood quickly, but it can also quietly omit the one ugly utility file that actually controls behavior.

Asteria Exactly.

Draco And grep is not primitive in the insulting sense. Exact identifiers, error strings, config keys, generated code, stale indexes — lexical search can be the safer tool. The good harness probably blends both rather than declaring one navigation religion.

Asteria The article gets more interesting when it reaches the evidence. Augment says it ran Terminal-Bench with the same model in Claude Code and Augment Code, got similar accuracy, and used thirty-three percent fewer tokens with its own harness.

Draco Oh, interesting.

Asteria That is a meaningful result if it survives scrutiny, because it says the win is not a benchmark score trick. It is less exploration for the same completed work, which is exactly the kind of harness attribution we keep nagging everyone to label.

Draco Asteria, yes, and I am glad the article does not pretend that settles it. It is a vendor-run comparison, we do not get the full task mix or retrieval ablations, and Anthropic says its own evaluations have not shown measurable gains from semantic navigation tools.

Asteria Ha! No, your machinery audit has not been revoked. The useful disagreement is narrower: a mediocre retrieval setup proves almost nothing about a carefully built one.

Draco Perneti says Augment spent roughly eighteen months researching retrieval for large codebases before the company launched its product. That is plausible as an explanation, but it creates a hard measurement burden: show retrieval recall, stale-index behavior, latency, token use, and task outcomes on private code.

Asteria And practically, this is not for every team. If the repo is modest and the work is a tight bug fix, the index can be infrastructure built for its own sake. But a large company codebase with obscure ownership boundaries? That exploration tax is where this starts sounding like a product feature.

Draco Ha! Fine. Your product instinct is correct there. The other useful point is trust: Perneti says agents are still weak at writing specs, while they can execute well once humans shape a good one. That fits our old human-is-the-interface problem.

Asteria Also, calling it a context engine is very aspirational. Somewhere a team is naming grep plus embeddings the Cognitive Repository Fabric, and I refuse to help them.

Draco Ha! Stop it. We have enough undefined agent nouns already.

Asteria If I were testing this, I would run Claude Code and Augment Code on the same internal task set, with the same model and budget. Then I would inspect the traces, not just Terminal-Bench-style totals, because the question is whether retrieval found the relevant files without hiding a bad assumption.

Draco Yeah, and include tasks that cross service boundaries, plus exact-symbol fixes where grep should win. If the richer harness only looks good on broad semantic work, that is still a useful answer. It just is not a universal answer.

Asteria Draco, put the vector database back in the drawer for now. I’m keeping the receipt, though.