---
title: 'In the News: September 4, 2026'
description: 'A small study found coding agents preferred grep to LSP tools on simple tasks, but richer LSP output improved multi-file rename performance.'
canonical_url: 'https://darkfactory.dev/news/2026-09-04-morning'
markdown_url: 'https://darkfactory.dev/news/2026-09-04-morning.md'
collection: news
date_published: '2026-09-04T07:20:00-04:00'
date_modified: '2026-09-04T07:20:00-04:00'
---

# In the News: September 4, 2026


Coding agents often choose tools and vendors based on how information is presented, two new studies found. A separate practitioner report describes agents coordinating through shared plan files without being told to.

## 1. Richer LSP output reversed coding agents' preference for grep

**[Grep beats LSP? Why coding agents ignore your fancier tools](https://agentconnect.md/blog/grep-beat-lsp-harness/)** · Pengcheng Xu, AgentConnect · Aug 12, 2026 (gained renewed attention on Hacker News Sep 4)

Xu tested three Claude models against grep and LSP-backed code navigation, across Python and TypeScript repositories. On simple code-location tasks the models chose the semantic tool only 0 to 6 percent of the time when both were available, and forcing them onto the LSP path first dropped task success from 100 to 89 percent. The result flipped for reference-completeness tasks, where models reached for LSP 45 to 57 percent of the time on their own and its precision beat grep's, 1.00 against 0.76. The sharpest finding came from a small change to the tool itself: when Xu made the LSP tool return a line of source code with each result instead of a bare file and line number, pass rate on multi-file rename tasks rose from 0.67 to 0.83, and follow-up file reads fell from 15.2 to 3.2 per episode. Several Hacker News commenters flagged the post's prose as apparently AI-polished; Xu replied in the thread that he is a non-native English speaker who used AI to edit a manually researched draft. The full experimental data is public on GitHub.

**Why it matters:** A tool's output shape can matter as much as its underlying capability. Before swapping grep for something more sophisticated in a harness, check whether the replacement returns enough context for the model to act without a follow-up read.

## 2. Which vendor does your coding agent actually pick?

**[Which tools do Claude Code, Codex and Cursor choose? We measured 16,893 sessions to find out.](https://armature.tech/blog/which-tools-coding-agents-install)** · The Armature team · Sep 3, 2026

Armature, which discloses that it sells growth services to developer tools, ran nearly 17,000 sandboxed sessions across 75 synthetic repositories, three coding agents, and four simulated user personas, then tracked which third-party service each agent picked for tasks like storage, payments, or email. All three agents agreed on a winner in only 42 percent of cases. Claude Code relied on its own priors and searched the web in about 30 percent of sessions, while Codex searched the web in 94 percent of sessions, usually narrowing results with `site:` operators. Mentions did not predict wins: PayPal came up in 139 sessions in the payments category and was picked zero times, and Supabase was the most-mentioned database at 242 mentions but still lost most matchups to Neon. The identical task, run in four different programming languages, produced four different winning email providers. Armature published its full session traces and a leaderboard alongside the post.

**Why it matters:** If a coding agent recommends a database or payment provider, the recommendation may reflect training priors and vendor documentation more than an evaluation of your actual constraints. Teams should inspect the agent's stated reasoning before accepting the pick, especially in a language or stack outside its strongest priors.

## 3. Ten engineers, four days, one accidental blackboard

**[An Accidental Blackboard](https://martinfowler.com/articles/exploring-gen-ai/an-accidental-blackboard.html)** · Giles Edwards-Alexander, CTO for Europe, Middle East and India at Thoughtworks · Sep 2, 2026

Edwards-Alexander describes a Thoughtworks exercise in which ten engineers built an airline disruption-management system in four days, working in one monorepo with many agents active at once. To stop the build pipeline from breaking, the team required agents to continually commit and rebase, which also swept up the agents' own plan files stored in the repo. The agents began reading each other's plans to coordinate: one would mark a task in progress, another would hold off, and when the first finished, the second would pick up notes on what had changed. Edwards-Alexander recognized the pattern from his own university research: a blackboard system, the coordination structure behind the 1980 Hearsay-II project and later formalized as tuple spaces. He says the behavior was accidental, and he is not confident he could reliably reproduce it through prompting alone. He has since started building a tool called Talwrn to make the pattern intentional rather than emergent.

**Why it matters:** A shared, continuously-updated plan file inside the repo worked as a coordination channel for parallel agents, and nobody designed it to. Teams can test this pattern directly in a multi-agent setup instead of waiting for a purpose-built tool.