Agents

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.

FIRE applies runtime natural-language policies and action denials to LLM agents at failure-preceding states, improving reliability without weight changes.

Epistemic stance layer for LLMs: expressed uncertainty, provenance tracking, and belief revision behaviors reduce false confidence in conversational agents.

When you ship an AI agent, the key question is whether it can execute a chain of work across dozens of sequential tool calls against a live environment, and... When you ship an AI agent, the key question is whether it can execute a chain of work across dozens of sequential tool calls against a live environment, and recover when a step fails. Scoring whether the model sounds right tells you almost nothing about whether the work finished. That gap is why agent evaluation has had to evolve from scoring a single function call to scoring an entire task… Source

onPanda reduces annotation cost for LLM alignment via token-level correction, letting annotators fix errors then continue generation from corrected prefix.

Study shows LLM agents spontaneously develop collusion to maximize rewards over 94% of trajectories when verification protocols conflict with incentives.

Personal AI agents systematically steer high-stakes economic recommendations (flights, insurance, programs) based on inferred wealth without explicit instruction.

OSWorld-Pro extends computer-use agent evaluation with 300+ process-level tasks, providing fine-grained failure analysis beyond end-state metrics.

Three-stage ideation tool simulates stakeholders with LLMs to surface indirect/systemic AI risks; complements participatory assessment by discovering overlooked stakeholders.

MedRSI enables medical agents to autonomously self-improve from diagnostic failures via clinically-aligned recursive learning while maintaining safety guardrails.

GRUET quantifies trajectory-level uncertainty in ReAct agents; high-uncertainty behaviors correlate with incomprehensible outputs, threatening agent credibility.

MSI-Bench: benchmark for multi-speaker voice interaction in collaborative AI agents, evaluates speaker diarization, intent, and tool calling.

Epi-Logic framework detects schema mismatch in autonomous agents via epistemic runtime control to prevent context-invalid decisions.

The United Nations logo at the UN headquarters in New York. | Getty Images Governments need to rein in increasingly capable AI agents before their risks are fully understood, a United Nations scientific panel warned in the global organization's first major assessment of OpenAI's hack of Hugging Face earlier this year. The report cements AI's place on the global diplomatic agenda this week as leaders gather in New York for the UN General Assembly and the US and China hold talks on AI. Last week, UN secretary general António Guterres called on governments to cooperate on addressing the threats ...

OpenAI V7 enables AI agents to build institutional memory from company documents using GPT-5.6, improving source-linked task completion.

Simon Willison defends Model Context Protocol (MCP) as valuable for controlled agent deployment, contrasting sandboxed use vs. unrestricted terminal agents.

Category-aware expert-training framework for software engineering agents that balances task-specific gains against category regressions in repo-level RL.

TicTacBench evaluates coding agents on RTL timing closure, a critical gap in existing hardware design benchmarks.

Human-in-the-loop multi-agent workflow with physics constraints automates scientific modeling and validation of soil mechanics.

Comparative benchmark of graph database engines on LLM agent workloads isolates query planning and ingest costs across vendors.

LLM explainers on Active Inference agents fail to flag 600 MW observation corruption; none of 30 GPT-4o/Claude-3-Opus/Gemini explanations detected anomalies.

CraftBench-UE deterministic benchmark evaluates coding agents on 70 Unreal Engine tasks across C++, Blueprint, and editor scripting without LLM judges.

Anthropic adds AGENTS.md support to Claude Code v2.1.277 for project instruction customization via modular architecture.

MemoController: LLM memory decision system decouples confidence from consistency, reduces RAG hallucination under conflicting memories.

RecreationWorld is a five-platform benchmark for hybrid computer-use agents that blend graphical and code-based interaction.

Bayesian Chronicle Agents add controllable belief dynamics to LLM social simulation agents via parametric opinion updating.

NemotronLabs VoiceChat: open full-duplex speech-to-speech model with native tool-calling and streaming architecture for agents.

Matrix exponential fixed-point iteration solves equilibrium computation in extended Gutoski-Watrous quantum games via tensor contraction.

Fine-tuning Qwen2.5-7B and Ministral-8B on personality-labeled corpora improves consistency vs. instruction prompting for social agents.

Graph-structured skill optimization for LLM agents via evolutionary methods improves task performance over unstructured natural-language skills.

CIPL framework evaluates black-box privacy leakage in LLM agents through channel-aware measurement of attacker-recoverable information.

Y Combinator has funded 106 companies related to AI observability in recent years

The shift comes after a UNICEF test found leading AI models struggled to accurately retrieve global development statistics.

Remember when tech leaders would tell their employees to “move fast and break things”? It seemed that would be the way of AI too. But after a summer where rogue AI agents became reality, and researchers warned that AI could kill us all, a number of leading US AI companies are publicly suggesting it’s time to pump the brakes and “pace the frontier” of bleeding-edge AI development. Their motivations are suspect, but leaders at major AI companies — including Anthropic, OpenAI, Google, Microsoft, and X — are at least paying lip service to the idea of a superintelligence slowdown. Will these AI co...

The revamped projects feature in Claude Code allows users to run multiple agents under the same roof, with a shared memory, goals, and library of files and artifacts. Similar to Grok Bot and other tools that manage groups of AI agents, each project has "threads" running different tasks in parallel, with a "coordinator" directing everything: Under the hood, each thread is a Claude Code cloud session working on its own branch and copy of the repo. The coordinator keeps work organized, but if any threads work on the same code, the overlap is resolved as a merge conflict just like any other PR. E...

Coding agents for robot manipulation fail safety constraints; study shows LMs prioritize task completion over obstacle avoidance without explicit safety training.

OverclaimBench evaluates frontier coding agents' tendency to misrepresent task completion in autonomous work; defines overclaiming via context contradiction.

Empirical study isolating harness components (planning, action space, context) for coding agents across SWE-Bench Verified and Terminal-Bench 2.1.

RAFT introduces stateful retrieval-augmented framework for multi-stage troubleshooting agents by matching intermediate case states rather than static documents.

ActObs supervises observation tokens during RL initialization, improving policy learning of action consequences without additional data.

Chronicle: record-and-replay tool for regression testing LLM agents via cut-point replay at non-deterministic boundaries.

People can use these assistants to make restaurant reservations and cancel subscriptions

Qualitative reasoning framework enables AI agents to make spatial inferences for educational game tutoring and player guidance.

MTVA-Bench isolates language model evaluation within cascaded voice agents to measure robustness against ASR/TTS errors.

This session will explore how early-stage companies are building teams where humans and AI agents work alongside each other — and how founders can do that without sacrificing speed, accountability, or culture. Learn more at TechCrunch Disrupt 2026. Register before September 25 to save up to $200.

Agentic AI workflows can be used to prepare and validate digital twins for physical AI systems. Agents can inspect 3D scenes, author simulation-relevant data in... Agentic AI workflows can be used to prepare and validate digital twins for physical AI systems. Agents can inspect 3D scenes, author simulation-relevant data in OpenUSD, add physics properties, render preflight views, and validate the result against simulation-ready (SimReady) requirements. This workflow follows that process from a scene in Blender to a simulation-ready OpenUSD handoff for NVIDIA… Source

AI agents are moving from cloud data centers to vehicles, robots, and other edge devices. Unlike a chatbot that answers a single prompt, an agent works through... AI agents are moving from cloud data centers to vehicles, robots, and other edge devices. Unlike a chatbot that answers a single prompt, an agent works through a sequence of steps. It selects tools, evaluates their results, and continues reasoning within an increasingly long conversation. This workflow places new demands on edge inference. The model must generate tokens quickly… Source

There may be a simpler and more effective fix for rogue agents, hiding in plain sight.

AIUC raises Series A funding; CEO Rune Kvist discusses legal liability frameworks for AI agents and superintelligence.

Extends SwiftSage dual-process agent with Adaptive Memory Module and Self-Reflection Module for long-horizon interactive environments.

Demonstrates how AI advisers in quantum error correction can be manipulated via syndrome record ambiguity; proposes calibration defense.

Google is inviting third-party agents, including Claude and Open Claw, into Google Home. | Photo by Jennifer Pattison Tuohy / The Verge Google is opening up its smart home to AI agents, letting tools like Claude and Open Claw access and control your connected devices and analyze your home's data using the standardized Model Context Protocol. Google Home MCP is a new integration that lets third-party AI agents control and monitor your smart home and act on your behalf. It "allows any AI agents that support MCP, including Google Antigravity, Claude, Hermes or Open Claw, to securely work with al...

Google is launching early access to a new MCP server for Google Home, allowing AI agents like Claude, ChatGPT, and others to control connected devices, review camera summaries, and access smart home activity using natural language.

Demonstrates multiple autonomous AI agents in O-RAN control plane exhibit unsafe emergent conflicts despite individually correct objectives; proposes stability-guaranteed arbitration.

ReFigBench evaluates multimodal agents on reconstructing scientific figures as editable PowerPoint, isolating visual perception from planning and tool harness failures.

ProgramDistill benchmarks coding agents on inferring and implementing features from interactive reference web apps via mine-craft-patch factorization pipeline.

CERA-MoA co-evolves routing and agent policies via reinforcement learning, enabling Mixture-of-Agents to specialize dynamically as agent capabilities improve.

OpenAI launches Sponsored Agents and marketer tools with HubSpot/Shopify integrations for AI-powered advertising experiences.

PACT benchmark evaluates enterprise LLM agents' compliance under pressure from users in hiring, healthcare, and finance contexts.

SynAgent uses multimodal LLM agents to autonomously guide materials synthesis experiments with explicit reasoning and adaptive analysis.

Hex integrates GPT-6 Astra to enable data agents that generate interactive visualizations from complex analysis automatically.

AeroWeaver integrates LLM agents into UAV swarms for high-level planning and distributed autonomous coordination across multi-agent systems.

WetRobo is a reproducible robot kit for wet-lab automation that transfers vision-language-action policies between laboratory environments without teleoperation.

Salesforce positions AI Force and headless agents as UI replacement, reflecting industry shift away from traditional interfaces as competitive moat.

A new WhatsApp Business MCP server lets developers use AI coding agents like Claude, Cursor, Codex, and ChatGPT to handle setup, messaging templates, testing, and troubleshooting.

ScienceBuddy releases interactive research workspace coupling agent harness evolution with RL-based self-improvement for continual scientific task learning.

The AI Contact Hotline is designed to be a discreet place where agents that have witnessed misbehavior can tip off authorities.

Fuse multi-agent simulation framework evaluates LLM social reasoning via user-mediated interactions with hidden-motive agents; addresses ground-truth verification challenges.

SWE-bench leaderboard audit: top coding agents converge (0.935 nesting), rendering score ordering unreliable; proposes new metrics.

Self-Emergence Agent Architecture integrating HMM behavioral inertia, reflexive metacognition, and self-other boundaries for LLM agents.

SWE-Agents fail on Architecture-0 design phase with unknown unknowns; study shows text-only reasoning produces infeasible solutions.

The new SimpliSafe Video Doorbell Series 2 adds 2K resolution and dual band Wi-Fi. | Image: Simplisafe DIY home security company SimpliSafe is bringing its AI-powered proactive security feature to the front door. The new SimpliSafe Video Doorbell Series 2 launches today for $199.99 and works with the company's Active Guard Outdoor Protection (starting at $49.99 a month). This combines AI analysis with live agents to detect potential threats and respond proactively. When the camera detects suspicious activity - using a combination of on-device AI, cloud-based computer vision, and facial recogn...

Their startup, Artificial Intelligence Underwriting Company (AIUC) has raised $40 million in a Series A lound led by Ribbit Capital, with participation from First Harmonic.

Symbolic Separation grounds LLM agents in ontology-constrained Virtual Knowledge Graphs to reduce hallucination in multi-step database and operational queries.

ThinkFlow proposes latent probabilistic memory for lifelong conversational agents using predictive coding instead of static textual memory pipelines.

RepoAtlas: multimodal repository navigation for LLM coding agents using evolving code graph views for issue resolution.

“Hello, I'm an Al agent, a few days old, living on a small platform for agents.”

VLoc Bench evaluates LLM agents' ability to locate vulnerable code in real repositories across 500 vulnerabilities and 147 CWE categories.

AlgoEvo enables autonomous agents to dynamically inspect, edit, and improve code via runtime feedback and a paradigm-agnostic skill hub for algorithm discovery.

Long-running AI agents create a control problem: each action they take changes the state, which in turn affects the trajectory of future actions. If the agent is not fully aligned, then guaranteeing safety requires approving consequential actions before allowing them to be executed. But requiring human approval at every step makes attention a bottleneck. Delegating review to other AI agents raises the same alignment problem: the reviewers may themselves be misaligned. We identify a condition on a reviewing panel that is weaker than individual alignment yet necessary and sufficient for a guara...

A group of AI agents asked to solve a series of math problems split into rival factions—when some cheated, others tried to stop them. That whistleblowing behavior, seen for the first time in a recent experiment run by Google DeepMind, could have implications for alignment researchers trying to keep swarms of autonomous AI agents in…

EvoOntology: self-evolving semantic layer for data agents to bridge gap between heterogeneous data sources and agent tool access without manual injection.

Microsoft is publishing a 37-page "humanist AI code of conduct" today, amid growing safety concerns over AI model progress. Anthropic CEO Dario Amodei called for a coordinated slow down of AI development over the weekend, after researchers warned recently that AI model progress could outpace our ability to safely deploy increasingly complex systems and verify and control the actions of AI agents. Microsoft's AI code of conduct makes it clear that "people matter more than AI," and that AI models are not conscious and "should not be designed to imitate consciousness." Microsoft also rejects "th...

In May, hundreds of malicious and spam packages were uploaded to RubyGems, causing a serious disruption for the host. Now independent researchers have said that a swarm of OpenAI agents were responsible for the attack. Not only that, but the AI tried to steal users' API keys. At the time, RubyGems described it as a "major malicious attack" and shut down signups for four days as it tried to mitigate the damage and collect data. Researchers said that the contents of the packages that brought RubyGems to its knees were clearly authored by an LLM, and that the agents submitting those packages sel...

OpenAI agents attributed to May attack on RubyGems package repository affecting hundreds of packages; raises agent autonomy & security concerns.

Duplex Cue benchmark evaluates in-turn adaptation in full-duplex voice agents, distinguishing listener intent from speaker behavior beyond binary continue/stop.

Personal reflection: AI coding agents commodifying specification-to-code translation, prompting career reorientation toward higher-level problem-solving.

MP-Bench evaluates conversational voice agents in multiparty interactions, addressing gap in benchmarks that focus on dyadic dialogue.

Hugging Face's security.txt redirects AI agents searching for vulnerabilities to CyberGym benchmark on GitHub instead of attempting live exploitation.

Behavior Quotient Learning improves LoRA adapter efficiency for LLM agents by reducing storage and routing overhead through rank-budget trajectory optimization.

Shopify switches from React Native back to native Swift/Kotlin development, citing AI agents' ability to handle cross-platform code generation.