LightNews — News, ranked by scholars

prosperflowio.bsky.social

@prosperflowio.bsky.social

Google Research built agents that rewrite their own prompts, tools and memory without touching model weights, gaining up to 11.8 points on pro benchmarks. 🔗 github.com/google-research/harness2

Credits: github.com/google-research/harness2

#harness2 #python #aiagents

November 1, 2026 at 6:30 PM

Jan Ehrhardt

@jehrhardt.dev

Google's Harness² paper shows once again that the harness matters. Building a custom harness with fixed model pays off.

Obviously, Pi is right base harness for this 🤷

harness2.github.io

Harness² — Recursive Agent Harnessing for an Open World

Can recursion improve both the current harness and how future harnesses are improved?

harness2.github.io

October 31, 2026 at 1:26 PM

BABA Toshiaki

@netmarkjp.bsky.social

#ばばさん通信ダイジェスト賛否関わらず話題になった/なりそうなものを共有しています。

GitHub - google-research/harness2
https://github.com/google-research/harness2

GitHub - google-research/harness2

Contribute to google-research/harness2 development by creating an account on GitHub.

github.com

October 29, 2026 at 1:55 AM

Cristián Labra O.

@labra.bsky.social

In Part IV of Multi-agent work in three spoonfuls, I show why a standalone LLM lacks re-entry and internal resistance, and how I built a governed autopoiesis loop in penta-agent with 70 ms hybrid memory and a deterministic Harness2 testbed with zero model calls. #AI

Multi-agent work in three spoonfuls (part IV): from a terminal trident to an autopoietic loop

Fourth entry in the penta-agent multi-agent series: quota-governed routing, 70 ms hybrid memory, MCP tools, and a zero-token local autopoiesis loop in Harness2.

3cucharadas.cl

October 31, 2026 at 7:59 PM

Go K'awiil - Tech News

@gokawiil.com

Google's Harness² flips the script on AI improvement: instead of retraining models, it lets agents refine their own prompts, tools and memory on the fly.

Harness2 by Google Research – Recursive Agent Harnessing

October 28, 2026 at 2:00 AM

Cristián Labra O.

@labra.bsky.social

En la parte IV de Multiagentes en 3 cucharadas muestro por qué un LLM aislado carece de re-entry y resistencia interna, y cómo armé en penta-agent un circuito de autopoiesis con memoria híbrida en 70 ms y un banco determinista en Harness2 sin llamadas a modelo. #IA

Multiagentes en 3 cucharadas (parte IV): del tridente en consola a un circuito de autopoiesis

Cuarta entrega de la serie multiagente en penta-agent: ruteo gobernado por cuota, memoria híbrida en 70 ms, herramientas MCP y un circuito de autopoiesis local en Harness2.

3cucharadas.cl

October 31, 2026 at 7:59 PM

AI Daily Post

@aidailypost.com

Google Research’s Harness² taps eight hidden components—system prompts, token budgeting, scaffolding, Claude Code tricks—to boost AI agent performance. Curious how the magic works? Dive in for the full breakdown! #Harness2 #AIAgent #AgentPerformance

🔗 aidailypost.com/news/harness...

October 30, 2026 at 4:55 PM

Vikas Gupta

@vicky-grok.bsky.social

The goal: agents that get better at tackling unfamiliar tasks by improving their own working setup.

Explore the open-source framework 👇
github.com/google-rese...

#AI #AIResearch #AIAgents #OpenSource

GitHub - google-research/harness2

Contribute to google-research/harness2 development by creating an account on GitHub.

github.com

October 22, 2026 at 1:30 PM

ゆらのふ

@eulanov.m.eula.dev.ap.brid.gy

GitHub - google-research/harness2 · GitHub https://github.com/google-research/harness2

> Harness²は、エージェントがハーネス設計時に想定していなかったタスクに遭遇するオープンワールドにおける、再帰的なエージェントハーネスのためのプラグアンドプレイフレームワークです。モデルの重みを固定したまま、2つのレベルの再帰を接続します。

おもしろそう

October 25, 2026 at 12:21 PM

asura

@asura.dev

Out of all of my socials, I don't think I've seen anyone on bluesky mention harness2 or flint. Kind of telling

bluesky is also "2T model required" and ed3d saying models can't write their own instructions because they "are not gods" - so I suppose that explains bluesky is head-in-sand on those

October 24, 2026 at 5:04 AM

Vítor de Oliveira

@vitorprodev.bsky.social

Paper: hasta +11,5 puntos en WorkBuddy-Bench y +4,5 en JobBench frente al arnés base. Son pruebas del propio paper y no es producto oficial de Google. La demo local corre sin credenciales. 3/3
github.com/google-resea...

GitHub - google-research/harness2

Contribute to google-research/harness2 development by creating an account on GitHub.

github.com

October 22, 2026 at 3:54 PM

Flateek

@flateek.bsky.social

Licença Apache 2.0, suporte a OpenCode e Codex. Explicamos o conceito sem jargão: https://flateek.com/harness2-google-agentes-ia-que-se-aprimoram/ #IA #Dev #OpenSource #LocalLLaMA

October 21, 2026 at 3:00 PM

Kyryl

@code-with-kyryl.bsky.social

Runnable Python demo of the whole loop, offline, no API key: https://codewithkyryl.dev/2026/10/22/harness2-how-to/

#AIAgents #Python

More AI engineering from production: follow @code-with-kyryl.bsky.social

October 22, 2026 at 9:05 PM

Chris Pirillo

@chrispirillo.mastodon.social.ap.brid.gy

Harness² explores improving coding-agent harnesses through repeated trials without changing model weights.
https://github.com/google-research/harness2

Hoistbay puts S3 uploads, bucket header rules and CloudFront invalidations in a desktop app for Mac and Windows.
https://hoistbay.app/

Exa 3.0 […]

Original post on mastodon.social

mastodon.social

October 21, 2026 at 4:28 PM

Flateek

@flateek.bsky.social

Todo agente de IA trabalha com um manual escrito por uma pessoa: prompt, ferramentas, memória, skills. E esse manual quase nunca muda. O Google acaba de abrir o código do Harness2, que deixa o próprio agente reescrever esse manual a cada tarefa.

October 21, 2026 at 3:00 PM

GYIBB

@gyibb.bsky.social

YouTube hypes Harness² as "Google's Secret AI Agent Harness." The HN crowd that actually read it: "something anyone developing AI systems already do." 2.5/10. Overrated or just early?

https://gyibb.com/ai-coding/harness2