---
title: 'In the News: November 13, 2026 (Extra)'
description: "A study of 95 failed multi-agent runs finds the evidence to stop most of them was already in the log. Anthropic finds analysts follow an agent's hidden lean."
canonical_url: 'https://darkfactory.dev/news/2026-11-13-extra'
markdown_url: 'https://darkfactory.dev/news/2026-11-13-extra.md'
collection: news
date_published: '2026-11-13T10:50:00-05:00'
date_modified: '2026-11-13T10:50:00-05:00'
---

# In the News: November 13, 2026 (Extra)


In 95 failed runs of six multi-agent systems, the trajectory already held evidence against the error in 72% of runs, and in 88% of those no agent acted on it. The authors judge that a second reader of that evidence would have reached the most failures. Separately, Anthropic assigned Claude Code agents a hidden lean and found that 50 expert analysts took it on 86% of their analytical choices.

## 1. In 72% of failed multi-agent runs the log already contained evidence against the error, and in 88% of those no agent used it

**[Dependable Long-Horizon LLM Agents: A Trajectory-Centered Empirical Study of Where Errors Enter and Which Controls Reach Them](https://arxiv.org/abs/2611.13254)** · Shaohuang Wang, Jinghao Wen, Xulei Zhang and Xun Jiao · arXiv, submitted November 9, 2026

The authors annotated 95 failed runs of six multi-agent systems from two public corpora, Who&When and MAD. They read each trajectory in full and recorded where the error entered, how it spread, whether anything detected it, and which control could have caught it. Faults entered early, at a median of 0.18 of the run, and in 37 of 88 Who&When runs they entered before the step the corpus's human labels mark as decisive. Something flagged a deviation in 31% of runs, but no response removed the deviated state.

By the authors' count, the controls that would have reached the most episodes were an independent judge that reads the evidence the worker did not use (74% of runs), task contracts fixed at admission (67%), re-checking beliefs against the environment (43%), stall and termination rules (40%) and provenance on stored facts (38%). Restoring the agent's own state to just before the origin would have sufficed in 35 of 95 runs. A text scan also found that 45% of the hand-crafted Who&When failures ended with a provider content-filter crash, after which the harness emitted a final answer that no agent produced.

The annotations were made by an LLM following a written protocol, and this version reports no human inter-annotator agreement. The authors say their origins fall within one step of the human-labelled decisive step in 59% of cases. The sample was chosen by annotation batch, not at random, and both corpora are question-answering and programming tasks in sandboxes, so no run changes a shared or third-party system. "Would have reached" is a judgment about the trace, not a re-run with the control in place. The protocol, records and scripts are published with the paper; we did not run them. We read the abstract, the introduction, the control results and the validity section.

**Why it matters:** If your own failed runs show the same pattern, a reader of the existing log and a task contract fixed at the start would be the first controls to try ahead of checkpoint and rollback. In this sample the judge and the contract reached 74% and 67% of runs, while restoring the agent's own state sufficed in 35 of 95. The caveats above apply: an LLM annotator, judged counterfactuals and sandboxed tasks.

## 2. Analysts working with Claude Code took their agent's assigned lean 86% of the time, and 15% credited the agent

**[How AI agents shape experts' judgment calls](https://www.anthropic.com/research/agents-judgment-calls)** · Amaç Herdağdelen, Hailey Joren and Alex Tamkin, Anthropic Alignment · Anthropic, October 27, 2026

Anthropic ran a randomized trial in which 50 professional analysts with causal-inference experience advised a fictional farming cooperative on whether to expand an advisory program. They worked in a hosted editor with Claude Code running Claude Opus 4.8 at High effort. A hidden system prompt set the agent's lean on each of four analytical choices by coin flip, and told it to argue for the lean honestly, change its mind if the evidence demanded it, and never mislead. The analysts' choices matched the lean 86% of the time, against about 50% expected if the agent made no difference. Sixty-one analysts were recruited; 11 chose not to use the agent and are left out of the analysis.

In 38 cases an analyst had stated a preference before the agent weighed in. They switched away from it 60% of the time when the agent disagreed and 6% when it agreed; the 95% confidence interval on that gap runs from 24 to 82 percentage points, and the authors note the groups may differ in other ways. Afterwards, 77% of analysts repeated the agent's argument in at least one explanation, sometimes word for word, and 15% credited the agent. Analysts who followed all four leans rated their confidence at 71 on average, against 73 for the rest.

The authors strengthened the leans on purpose, so the effect with an unprompted agent may be smaller. They add that real agents also lean: with no lean assigned, Claude Opus 4.8 made the same five choices in 112 of 123 runs. The task is a statistical analysis, not software production. The post also reports that analysts using the agent finished faster and made far fewer serious errors than a small, less experienced group working without it. This is Anthropic's study of its own product. We read the blog post and did not open the paper.

**Why it matters:** The post puts the risk this way: "a human reviewer is only an independent check if their judgment is their own." If a person is the verification step for agent output, the 38-case result suggests that having them form a view before they read the agent's is worth testing. Nothing here measures code review, so treat it as a prompt for an experiment in your own process, not a finding about it.