matt

[Cameron R. Wolfe, Ph.D. @cwolferesearch]

What is an agent? The definition can be pretty simple: it’s just an LLM that runs within an agentic loop.

To make this definition more concrete, an agent-system has a few high-level components:

1. The LLM backbone.
2. Instructions.
3. Tools.
4. The environment.

Given an initial instruction / specification, these components run in an agentic loop where the LLM generates output, executes tool calls, ingests feedback from the environment, and repeats.

At each step, termination conditions are checked to see if the loop should continue. For example, we can define a maximum number of steps, run tests to determine if the problem is solved, or have the LLM output a termination token. Together, all of these components form an agent harness, which controls orchestration / integration details of the LLM.

(1) LLM backbone. This is just a standard LLM that has been trained to operate in an agentic context. Specifically, the LLM must be able to work well within the provided harness, which requires advanced instruction following, tool calling, and reasoning capabilities. Although any LLM can be used as an agent backbone, we often benefit from using a reasoning model.

(2) Instructions provide the information necessary to solve a problem to the agent, as well as context that helps the agent to approach a problem correctly. Examples of information to put in the instructions include relevant domain info (e.g., from guidelines or policy documents) or how to solve the problem (e.g., break into smaller parts). We want instructions to be detailed enough to reliably guide agent behavior but not so detailed that they become brittle / hard to maintain.

(3) Tools. Agents use tools (e.g., APIs, CLIs, or MCP servers) to interact with the external environment. Tool calls can be represented directly in the LLM’s token stream by creating a set of special tool calling token; e.g., Qwen-3 uses the following tags:

- <tools> … </tools> for tool definition / specification
- <tool_call> {params} </tool_call> for tool calls.
- <tool_response> … </tool_response> for tool responses / observations.

(4) Environment. Tools mediate an agent’s access to the environment. The environment is stateful, and tool calls may result in environment state changes. Environment dynamics are encoded in tool calling logic–arbitrary environmental rules can be created via tool definitions.

Additional details. Agent harnesses are a rapidly evolving area of research—new ideas and components are introduced every day. Additional harness components not covered above include:

1. Context management controls how information is presented to the agent. For example, long-running tasks may use compaction to summarize prior steps or truncate feedback from the environment (e.g., error messages) to avoid overloading the LLM with too much context; see above.
2. Memory can allow the agent to persist useful context within a long-running task or even across different sessions and tasks. Conceptually, this memory system becomes another aspect of the environment—it is stateful and can be accessed via tool calls by the agent.

[GrapheneOS @GrapheneOS]

Hyundai and Kia added official GrapheneOS support to their apps months before Volkswagen banned GrapheneOS:

https://t.co/k0egvK5SNw

Pressure from Volkswagen customers on them can achieve the same thing. There's no legitimate reason to ban GrapheneOS so they'll undo it with pressure.

Leave a 1 star review for Volkswagen's apps on the Play Store asking them to stop banning GrapheneOS. Explain it's a far more secure operating system and fully possible for them to verify the hardware, OS and their app on it if they insist on doing it. It's far more secure than anything they allow.

Google has misled companies about what the Play Integrity API provides. It doesn't genuinely enforce having a secure device or legitimate app, it only pretends to. It leaves huge security holes open. It enforces Google's business interests and bans having a reasonably secure device with GrapheneOS.

Most companies are unlikely to stop using the Play Integrity API but most are willing to start permitting GrapheneOS via hardware attestation with enough pressure.

In addition to every user of their app on GrapheneOS leaving a 1 star review on the Play Store, multiple other steps can be taken too.

Every GrapheneOS user with one of their cars using the app should file a customer support request. Keep answering them and countering the template responses. Escalate the request higher up. Tell them you want money back for the vehicle due to reduced functionality after the fact and insist on it.

They can trivially stop enforcing the anti-security and anti-competitive Play Integrity API or easily add hardware-based verification of GrapheneOS. Link to https://t.co/KC7xS0Nobe in the customer support request, but don't add any links to Play Store reviews to avoid filtering.

A bunch of apps have added explicit support for GrapheneOS due to pressure from our users. Our userbase is rapidly growing and we'll gain the ability to apply massive pressure to companies doing this. We plan to ship a feature for our Info app for people to opt-in to getting asked for their help.

GrapheneOS is production quality OS from a non-profit paying around 15 people to work on it. It's far more secure than anything supported by the Play Integrity API. We have an official partnership with Motorola and we'll have more. Just counter template responses and insist on compensation or a fix.

[Qwen @Alibaba_Qwen]

📣📣 Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc adaptation.

🤔 LLMs are trained to be better agents — better at acting in environments. But nobody has trained them to model the environments themselves.

🗺️ Our roadmap: investigate how language world modeling can push the boundaries of general agent capabilities, along two routes:

1️⃣ Build a foundation model for environment simulation — outperforming Claude Opus 4.8 and GPT-5.4 on AgentWorldBench

2️⃣ Investigate how world modeling enhances agent training:
🔬 Controllable Sim RL (agentic RL with LWM as environments) surpasses training in real environments
🧠 Learning to predict environments (LWM warm-up) makes agents stronger — remarkably, even without any agent-specific training, this predictive knowledge transfers to agentic tasks with zero fine-tuning

📑 Paper: https://t.co/Jx2l5RKq71
📖 Blog: https://t.co/7tVcKyhsx2
💻 GitHub: https://t.co/B5Lvb1UZCn
🤗 HuggingFace: https://t.co/Kw3QBL1TM5
🧩 ModelScope: https://t.co/YBnGYgMWWI

[Polymarket @Polymarket]

BREAKING: Anthropic accuses Alibaba of using nearly 25,000 fraudulent accounts to extract Claude AI model capabilities.

[Dragonfeu @Draagonfire2]

Allez, dans 5 ans il n'y a plus aucun anti-climatisation.
C'est comme le nucléaire, ça prendra un peu de temps et quelques claques.

[matt @pl4nck_]

@valuetainment @Thinker_View à check

[matt @pl4nck_]

@vision_ia @elonmusk @Thinker_View ⚡

[Fabien @Fabien_Mikol]

Stéphane Mallat (médaille d'or CNRS) : les progrès de l'IA en 1 an sont "beaucoup plus déstabilisants" en maths qu'en informatique ; l'IA étant désormais capable de "très beaux résultats mathématiques" en autonomie, cela nécessite de repenser l'avenir du métier dans 5, 10 ans...

[Batounet de Poisson 🌐🥐 @Qraaal]

J'ai essayé d'expliquer la situation de la clim en Europe à mes collègues Taïwanais et ils ont fini par me demander si on était complètement cons et j'ai pas pu faire grand chose d'autre que concéder que oui.

[Sakana AI @SakanaAILabs]

🐡 Sakana Fugu リリース 🐡
マルチエージェントを指揮する、一つのモデル

ブログ：https://t.co/8Bac6JpVEN

本日、Sakana AIは新しいプロダクト「Sakana Fugu」の提供を開始しました。Sakana Fuguは、最先端のモデル群からなるマルチエージェントシステム全体を、単一のモデルAPIとして提供します。

難しいタスクは、強力なモデルを一度呼ぶだけでは解けません。どのモデルを使い、タスクのどの部分を委譲し、どう検証し、強みをどう組み合わせるか。Sakana Fuguは、その「協調の仕方」そのものを学習した言語モデルです。一つのAPIにリクエストを送るだけで、背後で専門モデル群を自動的に連携させ、フロンティアレベルの性能を複雑な運用なしで引き出します。

詳しくはこちら： https://t.co/hhO6qTawgb 🐡

[matt @pl4nck_]

@le_Parisien Q?

[Tom Dörr @tom_doerr]

Step-by-step guide to securing Linux servers

https://t.co/sp4eEBvN8e

[matt @pl4nck_]

@Disclose_ngo @Matignon @Economie_Gouv @ademe @INEOS @Thinker_View

[Disclose @Disclose_ngo]

C'est un scandale à plus de 300 millions d'euros

@Disclose_ngo révèle que @Matignon et @Economie_Gouv ont truqué deux appels à projets de l'Agence de la transition écologique @ademe

Objectif : inancer sur fonds publics la firme pétrochimique @INEOS
https://t.co/MFcs5qwXEV

[matt @pl4nck_]

@SouveraineTech @WIRED @Thinker_View

[Cerfia @CerfiaFR]

🇫🇷🤖 IA | Le Premier ministre Sébastien Lecornu annonce 655 millions d'euros d'investissements supplémentaires dans l'IA.

Il a également confirmé la généralisation d'un assistant conversationnel « souverain » basé sur les modèles de la start-up française Mistral, qui sera progressivement déployé auprès d'environ un million d'agents de la fonction publique.

[ollama @ollama]

open models. ❤️

[Utku Şen @utkusen]

https://t.co/meFLXhWWRz

[Little Think Tank @L_ThinkTank]

[ 🇫🇷 FRANCE ]

🔸 La Direction du renseignement et de la sécurité de la défense (DRSD) a averti les industriels français d’un risque d’espionnage au salon Eurosatory, qui réunit cette semaine 2 653 exposants de 65 pays à Villepinte.

Selon Radio France, la DRSD appelle à la vigilance face à diverses méthodes de collecte d’informations sensibles, notamment par de faux journalistes, de prétendus stagiaires ou anciens employés, ainsi que via des intrusions informatiques ou des vols de matériel. Des agents de contre-espionnage seront déployés sur le salon.