---
title: 'In the News: November 5, 2026 (Extra 3)'
description: 'A Berkeley paper has an agent rewrite its own harness mid-task, and Rivet launches an actor runtime as an alternative to per-agent sandboxes.'
canonical_url: 'https://darkfactory.dev/news/2026-11-05-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-05-extra-3.md'
collection: news
date_published: '2026-11-05T23:20:00-05:00'
date_modified: '2026-11-05T23:20:00-05:00'
---

# In the News: November 5, 2026 (Extra 3)


Two items on how agent harnesses and the runtimes under them get built. A Berkeley-led paper reports lower model cost when an agent edits its own harness during the task, and Rivet launched an open-source runtime that gives each agent a lightweight actor instead of a sandbox.

## 1. A Berkeley paper has agents rewrite their own harness during the task and reports lower model cost

**[Exo: An Architecture for Self-Improving Systems at Runtime](https://arxiv.org/abs/2611.00463)** · Alexander Krentsel, Sylvia Ratnasamy, Ion Stoica (UC Berkeley), Martin Casado (Andreessen Horowitz), Ankur Goyal (Braintrust) · arXiv, October 30, 2026

The authors describe Exo, a runtime in which the agent revises its own harness policy (context construction, tools, memory) while it works, using the same code-editing tools it uses for ordinary tasks. A small protected layer keeps execution history and controls when a new policy goes live. The authors report that on long-running objectives in an emulated RuneScape environment, Exo cut model cost by 12 to 53 percent against Codex and Pi in standalone runs and by 45 to 82 percent in transfer runs, "while matching task completion rates." Against the same starting policy without the ability to edit itself, they report 35 to 61 percent lower held-out cost on AppWorld and SREGym with comparable correctness. On Continual Learning Bench coding tasks, cost including learning fell 14 percent, with both policies resolving every issue.

These are the authors' own results on a game environment and two benchmarks, with models the paper names as gpt-6.1-sol and gpt-5.6-sol. The coding result is the smallest of the three. We read the evaluation sections; we did not find a code release or an independent replication.

**Why it matters:** Most harness tuning today is a human editing prompts and tool lists between runs. This paper moves that edit inside the run and keeps a protected layer that decides what activates. If the result holds, the audit trail of an agent's self-edits becomes something a team would need to review. The paper's transfer runs, where the agent reuses a harness it already specialized, account for the largest savings it reports.

## 2. Rivet launches an open-source actor runtime for agents, with vendor benchmarks against sandboxes

**[Launch HN: Rivet (YC W23): Scale and Orchestrate Agents on S3](https://news.ycombinator.com/item?id=50381687)** · Nathan Flurry and a co-founder, Rivet · Hacker News, November 5, 2026

Rivet's founders say each agent gets a long-lived actor with in-memory state, durable SQLite storage, and hibernation when idle, instead of its own sandbox. Writes are acknowledged once SSDs in two of three availability zones hold them, then batched to S3 in the background. In their [published benchmark](https://rivet.dev/docs/benchmarking/), run November 5 on Rivet 3.0.3 in Azure, a new actor cold-starts in 4.30 ms at the median (4.99 ms at p99), and a durable write takes 1.07 ms at p99. The founders say Amp moved its multiplayer agents to Rivet from Cloudflare Durable Objects; that is their account, and we did not verify it.

The headline comparison to sandboxes is weaker than the Rivet numbers. Rivet's own page says the E2B figures (1.27 s median cold start, 512 MiB memory) come from E2B's documentation and a third-party benchmark, not from the same setup, and calls them "separate benchmarks with different concurrency and environments." The 10x storage saving is a calculation from AWS list prices that excludes read and write costs. The page says the repository has scripts to reproduce the Rivet numbers; we did not run them.

**Why it matters:** The design question is whether an agent needs a full machine while it waits on a model. For agents that mostly call tools and APIs, an actor may be enough. A coding agent that has to run a compiler and tests still needs a process to run them in, and one commenter asked Rivet that exact question in the thread.