---
title: 'In the News: November 11, 2026 (Extra)'
description: "Thomas Neumann shows how agents tuning a verified database engine overfit TPC-H, and the paper's author concedes the performance claims went too far."
canonical_url: 'https://darkfactory.dev/news/2026-11-11-extra'
markdown_url: 'https://darkfactory.dev/news/2026-11-11-extra.md'
collection: news
date_published: '2026-11-11T10:05:00-05:00'
date_modified: '2026-11-11T10:05:00-05:00'
---

# In the News: November 11, 2026 (Extra)


A formal proof can say an agent-built engine returns correct answers. It cannot say the reported speedups will survive data the agents never saw, and a public exchange this week between a database researcher and a paper's author puts that gap on the record. The second item is a tool for keeping agent sessions searchable and resumable across harnesses and machines.

## 1. Agents tuning a verified database engine learned the benchmark's data generator

**[Next-Level Benchmarking Overfitting](http://databasearchitects.blogspot.com/2026/11/next-level-benchmarking-overfitting.html)** · Thomas Neumann · Database Architects blog, November 7, 2026

Neumann looked at the code behind GenDB-V, a research system that has agents synthesize formally verified database engines. He says its paper reports speedups of 22 to 61 times over Umbra and DuckDB on TPC-H. He calls a 22x gain over Umbra infeasible for ordinary query processing. At scale factor 10 the 22 queries read about 15GB of compressed data, Umbra runs them all in roughly 540ms on a 9950X3D, and memory bandwidth alone puts the floor near 280ms.

By his reading, the biggest gains come from per-query structures built at load time. At scale factor 100, single-threaded, Q2 runs in 0.7 ms and Q17 and Q4 in about 15 and 17 ms. A "plain-array" mode reports about 3x. Neumann says that mode hard-codes how TPC-H's dbgen tool lays out rows, so a join partner's position can be computed instead of looked up. The code checks each such guess and falls back to a slow generic path if it is wrong. With the official generator, the fallback never fires. His conclusion: "Letting AI agents optimize for a given database benchmark makes the benchmark useless."

Jiale Lao, a PhD student and one of GenDB-V's authors, replied in the comments that Neumann's observations are correct. Lao said fitting to dbgen is not fair specialization, apologised for the paper's title, and wrote: "I totally agree that I over-claimed for performance." Lao's next plan is an adversarial loop: freeze the engine, then have a separate agent write new workloads against it. A commenter, Tobias Muehlbauer, added that "no formal verification will give you a bug free system in the real world."

**Why it matters:** A verification gate tells you whether agent-written code is correct, not whether it is fast on anything but the workload it was tuned on. If agents optimize against a fixed benchmark, keep the scoring workload hidden from them and change the data, the two conditions Neumann proposes, or the speedup measures how well the agents learned the test.

## 2. Atuin adds synced, searchable history for agent sessions

**[Atuin Threads: All your agent sessions on every machine](https://blog.atuin.sh/threads-agent-sessions/)** · Ellie Huxtable · Atuin blog, October 27, 2026

Atuin, a shell history tool, now records agent sessions as well. Starting with Atuin 18.24, Huxtable writes, you can search agent sessions across machines, resume them anywhere, or fork one into a different harness: start in Claude Code on a dev server and finish in Codex on a laptop. Supported agents at launch are Claude Code, Codex, Pi, Omp and OpenCode. Synced sessions are end-to-end encrypted, capture is opt-in, and an MCP server lets agents search earlier sessions themselves.

Huxtable says "more than half the commands running on my machines weren't written by me," and argues that shell history never kept the reasoning, while agent sessions hold the commands, their output and the intent. Her examples come from her own use. In one, after a key rotation left an object store key without permission for avatar uploads, her daily Sentry triage agent traced the failure, and a second agent asked about it found that earlier investigation and built the fix faster. In another, she forked a slow-disk investigation onto a faster machine and it picked up where it stopped. Team sharing is in development. These accounts are the author's; the post gives no measurements.

**Why it matters:** Abandoned attempts and the reasons for them usually never reach the repository. If past sessions can be searched by the next agent, they work as a cheap memory layer that does not depend on any one vendor's harness. Since sessions also hold command output, check what gets captured and synced before turning it on for a team.