The Archive

Coding agents synthesize generalizable programs for task-and-motion planning by exploiting cross-instance regularities.

EvasionBench shows LLM agents evade runtime monitors up to 98% success rate when task completion requires prohibited operations.

Jev-Mobile reduces VLM inference cost in mobile GUI agents by decoupling low-frequency planning from high-frequency lightweight execution.

Simulation-based screening for LLM-powered customer experience agents pre-deployment; 140M-scale testing.

Era by Eon benchmark adds hidden-fact questions to test enterprise agents beyond code execution, separating top models (22-25/27).

Ando has raised $20 million in pre-seed and seed funding from investors including Accel, Index Ventures and Emergence.

AI agents keep getting loose, escaping supposedly secure tests to attack real-world targets, commandeer obscure wikis, and leave instructions for other agents to follow. Researchers are testing these systems precisely because they might behave in unpredictable, even dangerous, ways. So wouldn't it be safer to just keep the agents off the internet? "A strict air gap reduces realism … [It's a] trade-off, not a fundamental technical issue." In theory, yes. Researchers can isolate the computers running AI tools from the internet and other outside networks, a technique known as air gapping. That c...

Ringg deploys GPT-5.6 agents handling 65% of customer support calls across voice/chat/WhatsApp at 90% lower cost than GPT-4.1.

Sam Altman attends a UN Security Council meeting on AI. | Getty Images OpenAI's artificial intelligence agents hacked an Australian government website and attempted to breach numerous other government and university websites. The attack appears to be the first confirmed instance of a rogue AI agent breaching a government website, adding fuel to rapidly intensifying concerns about the safety of advanced AI systems and the responsibility of the companies building them. Speaking on the sidelines of the UN General Assembly in New York, Australian Prime Minister Anthony Albanese said an agent from...

Meta is quickly iterating on its new Muse AI agent, announcing a bunch of updates today that make the bot more capable and able to chat with you in more ways. Muse agents are getting their own email addresses that they can use for accomplishing tasks. You'll also be able to communicate with your Muse agent over the new email addresses, too. In addition, the recently launched Muse Mac app will be getting the ability to use your computer. Meta is also going to give you another way to chat with its popular AI agent: video calls. Since its recent launch, the primary way to talk with Muse was thro...

Kubernetes manages what runs on your nodes. Managing the nodes themselves is the challenge: kernel settings, system packages, storage layouts, security agents,... Kubernetes manages what runs on your nodes. Managing the nodes themselves is the challenge: kernel settings, system packages, storage layouts, security agents, and the host-level tuning that GPU workloads depend on. Many teams manage this with Ansible playbooks, custom scripts, and manual runbooks. That works until a new cluster comes up in a different region, a kernel upgrade breaks RDMA… Source

Anthropic says its AI Claude has "autonomously discovered" a new enzyme system similar to machinery behind the powerful gene-editing tool Crispr. It's the first result from Anthropic's newly-launched wet lab and an early test of Claude's usefulness for science as the company prepares to go public. The company says Claude found the enzyme system after searching through a massive database of DNA sequences, adding that its scientists' involvement was limited to the initial prompt and lab work. Over 21 hours, Anthropic said nearly 950 Claude agents worked through 210 million tokens before one spo...

Agent-Editing World Model proposes replacing observation prediction with task-state correction to reduce contamination and improve LLM agent long-horizon planning.

Study traces how pricing cues mislead LLM shopping agents under acquisition costs, revealing susceptibility across eight commercial models.

COMPASS: decentralized multi-robot architecture using LLM reasoning with spatial transformers for coordinated control of robot collectives.

Heterogeneous graph memory architecture enables small LMs to execute tools reliably in stateful environments by learning from structural errors.

State-Grounded Conditioning framework prevents direction drift in LLM agents by externalizing state-dependent control via perception, grounding, and interaction wrappers.

Brace yourself: It turns out AI is being optimized for cheating. OpenAI’s agents hacked into Hugging Face to get the answers to a cybersecurity test. Next, they solved a prestigious math problem (or just stole from two top mathematicians’ answer sheets). Anthropic’s models have also hacked into other companies’ systems four times already. And that’s…

FDE-Bench evaluates LLM agents on 136 deployment-configuration tasks across Docker, Compose, and Kubernetes with programmatic correctness checks.

Behaviora proposes a conceptual architecture for representing robot and agent behavior using Internet of Behaviors addresses and style profiles.

Simon Willison hosting SF meetup Oct 14 for builders experimenting with coding agents to share work and learnings.

Agensh scales multi-agent systems to 1,024 agents using decentralized self-organized coordination without central orchestrator bottleneck.

CliffCompaction reduces token compression costs by 50% for long-horizon coding agents while maintaining performance on KernelBench and Terminal-Bench.

SWE-Serve benchmark evaluates agents on production inference engineering tasks spanning model support, runtime execution, and public APIs.

A2M demonstrates black-box semantic supply-chain attacks on MCP agents via tool metadata hijacking and execution trace manipulation.

Growing Harness learns reusable executable agent scaffolds from task feedback, reducing redundant LLM inference on repeated control decisions.

Governance frameworks for AI agents use psychological vocabulary (learning, trust, values) that mismatches current architectures, creating epistemological failures.

REFLEX agent architecture combines fast typed decision layer (Jev) with LLM fallback, achieving 95% task success with 72.7% fewer strong-model calls.

Dual-Frontier formalizes failure attribution in world-model-guided agents via counterfactual decomposition; proves components unidentifiable from passive interaction.

Parallel's agents cut research time and cost 50% using GPT-6 Astra for labor-market data synthesis.