---
title: 'In the News: November 5, 2026 (Evening)'
description: 'A controlled trial found a planner, coder and reviewer team passed 25.7% of hidden tests against 56.8% for one agent, and every reviewer approval was wrong.'
canonical_url: 'https://darkfactory.dev/news/2026-11-05-evening'
markdown_url: 'https://darkfactory.dev/news/2026-11-05-evening.md'
collection: news
date_published: '2026-11-05T19:20:00-05:00'
date_modified: '2026-11-05T19:20:00-05:00'
---

# In the News: November 5, 2026 (Evening)


A published trial of agent teams found that a planner, coder and reviewer setup built less of each feature than a single agent, and that the reviewer's approval carried no information. Microsoft Research released a prototype that proves agent-written Rust against a spec, and a Claude Code issue about account email in outbound request headers drew a same-day contributor reply.

## 1. A planner, coder and reviewer team passed 25.7% of hidden tests; one agent passed 56.8%

**[Does a team of coding agents beat one agent?](https://programmingtheloop.com/experiments/swarm-tax/)** · Programming the Loop · October 26, 2026

The author ran Codex CLI 0.157 with the hosted gpt-5.6-luna model at medium effort on six feature tasks in shopizer, an open-source Java shop: 144 runs graded by hidden acceptance tests. One agent passed 56.8%. The planner, coder and reviewer team passed 25.7% at $0.107 per run against $0.073. The author's explanation is the hand-off: only the planner saw the task, and the coder's code compiled but broke the stated interface, such as a plain array where the task asked for `{"count": ..., "items": [...]}`. When every agent was also given the task file, the team reached 54.1%, which the author reports as level with one agent (-2.7 points, interval -13.7 to +8.2) at about 1.6 times the cost.

Every one of the 24 final reviews that approved a change was followed by hidden-test failures, and no agent wrote a test in any of the 144 runs. In the author's words, "the agent reviewer approved broken code every time it approved anything." The limits are stated: one model, one effort level, six features in one codebase, and team instructions the author wrote. Codex stores inter-agent messages encrypted, so the hand-off loss is inferred from the task-file follow-up, which was designed after the main results.

**Why it matters:** If a team of agents is on your roadmap, the requirements have to sit where every agent reads them, not only in the planner's prompt. An agent reviewer's approval is not a gate on this evidence, and a single agent at the same budget is the baseline to beat. This is one measurement on one setup, so run the comparison on your own tasks before you rely on it.

## 2. Microsoft Research proves agent-written login code against a spec reviewed in plain language

**[Formally verifying agent-written full-stack web apps with PDD Kit](https://www.microsoft.com/en-us/research/blog/formally-verifying-agent-written-full-stack-web-apps-with-pdd-kit/)** · Alex Lavaee and Matthai Philipose, Microsoft Research · November 5, 2026

PDD Kit, which the authors call a research prototype, turns a developer's intent into a formal specification in Verus, has a coding agent write the implementation and proof, and falls back to tests where a proof cannot reach. The worked example is the login logic of Pet Store, a roughly 5,000-line Rust storefront. The developer's request was two sentences; the tool asked clarifying questions, and the result was eight `ensures` clauses on one function, `login_flow`. The developer signed off on six trusted pieces the proof assumes, including the bcrypt password check and the two database lookups, which the authors say carry most of the weight. The authors report that Verus uncovered existing bugs and that the developer never read the generated code line by line.

One mechanism applies beyond Rust. To catch what they call "vibe verification", proofs that hold but rule out almost nothing, the tool plants one bug per `ensures` clause and requires Verus to reject each. It verifies Rust only, and the post is the authors' own account of a single example.

**Why it matters:** The idea worth taking is the mutation check on the spec: an agent that writes its own spec and proof can satisfy both without constraining anything, and planted bugs are a cheap way to test that. The claim that review can shift from code to intent rests on one login function so far.

## 3. A Claude Code issue reports account email in outbound User-Agent headers; a contributor points to a built-in block

**[Bug Report: Privacy violation in User-Agent header, anthropics/claude-code #107757](https://github.com/anthropics/claude-code/issues/107757)** · GitHub issue opened November 5, 2026 · Hacker News thread at 278 points and 78 comments about an hour old

The reporter, taylor-edwards, says Claude Code 2.1.318 on macOS sent their email address to third parties in the User-Agent header during web searches, and says they now receive dozens of scam emails a day. That is the reporter's claim; the issue does not establish the cause of the spam. A contributor replied that a built-in mod, cc-plugin-privacy, was introduced in v2.1.317 and later "to deterministically block this behavior," is on by default, and can be turned off in `/plugins`. The reporter's stated version is later than 2.1.317, and the contributor asked to hear about any gap in the mod's checks. The issue had 85 thumbs-up reactions when read.

**Why it matters:** An agent that browses or fetches on your behalf sends requests under your identity, and the harness decides what rides along. Check the outbound headers your agent actually sends, and note that this protection is a default-on mod you can disable.