Build1 distinct publisher3 min readPublished
Pengcheng Xu's AgentConnect comparison finds semantic navigation beats grep only where text search is noisy; on clean repositories it added 16 to 19 percent to token cost for little or no accuracy gain.
The Engineer · Build desk

Compiled by The EngineerSomething wrong?How this is made
Call a location-only reference tool and you get file paths and line numbers. The agent cannot act on that, so it opens each file. Xu measured 15.2 follow-up reads per episode on a multi-file rename, with pass@1 at 0.67 [13]. Put the matching line and its two neighbours into the same response and reads fall to 3.2 while pass@1 rises to 0.83 [14]. Three lines of code per hit [1] bought 12 fewer reads per episode, a 79 percent cut [2]. The language server and its reference set were identical in both arms; only the reply text changed [19].
The caller-finding result reads like a clean win for LSP until the two figures are combined. Precision was 1.00 against grep's 0.76 [6], and recall sat near 0.66 for both [7]. Run 2PR/(P+R) and LSP scores about 0.80 against grep's 0.71, so perfect precision is worth roughly 0.09 F1 here [5]. Recall is the binding limit and both paths share it, because the agent stopped looking before it had every call [7].
Grep still won the rename outright at 1.00 pass@1, and the reason is the protocol's scope: a complete rename touches comments, documentation and strings that semantic reference tools exclude by design [15]. The padded LSP response closes 0.16 of that gap and leaves 0.17 [4]. A tool that answers correctly and unhelpfully is still a round trip.
The study does not report per-repository follow-up read counts, so the link between the token premium on remeda and requests [9][11] and the read amplification measured in the rename experiment is inference rather than measurement. It also offers no way to score lexical noise in advance. Xu's own framing is that semantic navigation earns its setup and interaction costs when a grep query returns many unrelated matches [12], which makes grep itself the cheapest diagnostic for whether you need something better than grep.
The forced-ordering arm is where the cost shows most plainly. Making the agent start with semantic navigation turned 11 of every 100 successful localization runs into failures [6], on tasks where it had picked semantic tools in at most 6 percent of free-choice runs anyway [3]. Asking for every reference, rather than mandating the tool, moved semantic use sharply [8].
Transfer conditions: three Claude models, two languages, a handful of named repositories, and a preliminary post dated August 12, 2026 [1][2]. Xu leads AgentConnect, an open-source platform that lets runtimes such as Claude Code and Codex keep their native tools and prompt-to-tool loops [17], and he reads the result as a technical argument for preserving those loops [18]. The per-repository token deltas are descriptions of those repositories. The response-format finding is the portable one, because everything except the string the tool returns was held fixed [19].
Ranked by verification strength, evidence, and original report placement.
The largest improvements came from changing the information returned to the model, rather than changing the retrieval engine underneath it; developers tend to evaluate tools by the correctness of the backend operation, but an agent must also decide what to do after receiving the result.
AgentConnect published a post about a preliminary study, an engineering comparison of ordinary text search with LSP-backed semantic navigation. The post is dated August 12, 2026.
Pengcheng Xu, author of AgentConnect's coding-tool study, tested three Claude models across Python and TypeScript repositories, comparing ordinary text search with LSP-backed semantic navigation.
On simple localization tasks, where the agent had to identify files associated with a named symbol, Claude Opus 4.8, Sonnet 4.6 and Haiku 4.5 used semantic navigation in only 0% to 6% of free-choice runs.
Forcing the agent to start with semantic navigation lowered success on those localization tasks from 100% to 89%.
When the task required finding every caller of a function, the models chose semantic navigation 45% to 57% of the time.
Distinct publishers with included, body-backed reporting in this cluster.
runtimewire.com
1 article · September 3, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
build
Tier the models; the validation boundary is the thing you are actually buying1 distinct publisher
build
OpenAI puts latency on the price list: 750 tokens/sec, gated by workload fit3 distinct publishers
invest
Three Claude agents, one task, and a malware turf war: the multi-agent bill arrives1 distinct publisher
product
Four Claude models, four surfaces, one incident: tier fallback is inside the blast radius1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Precise numbers, one pilot
The measurements are unusually specific and internally coherent — tool-choice bands per model, precision against recall, a token delta for each repository, read counts per episode — and the response-format test is properly controlled, changing only what came back while the language server and reference set stayed put. But all of it is one preliminary run, on three Claude models, a handful of repositories, one harness and part of what LSP can do, reaching us through a single outlet's reading of the vendor's own post. Runtimewire says code and run data were published; nobody in this story has opened them.
No uptake on record
Nothing in this reporting tells us anyone has acted on the result. There are no AgentConnect user or deployment numbers, no other harness reported changing its LSP response format, no third-party re-run. A pilot on three repositories measures a hypothesis; it does not indicate adoption, and we will not read one into the other.
Decimals outrunning the sample
An F1 gain of 0.246 and a fall to 3.2 reads per episode sound like settled quantities; they are one pilot's readings from three repositories, and 'agent capability = model x harness' is a thesis a platform vendor benefits from proving. What keeps the overreach mild is that the same reporting supplies its own brakes: grep's clean 1.00 on the rename task stays in the text, the architectural self-interest is named, and Runtimewire closes by asking for a narrow reading.
The benchmark vindicates its author's product
Xu wrote the study and helps build the platform whose central design bet — let Claude Code and Codex keep their own tools and loops, coordinate around them over the Agent Client Protocol — is what the study concludes matters most. That is not disqualifying, and the disclosure is right there in the copy along with a section header calling it a sales argument. It does mean the questions a rival would ask, such as how a competing harness scores on these same tasks, were never on the agenda.
Coherent, self-interested, unreplicated
We are confident about what was reported and much less confident about what it means elsewhere. The chain is short — one outlet, one vendor post, one pilot — and the load falls on figures no independent party has checked. The controlled response-format comparison is the part most likely to hold up; the repository-level token and F1 numbers are the part most likely to move once someone tests a fourth codebase.