The Archive

FutureSim: Benchmark measuring frontier agents' ability to adapt and forecast beyond knowledge cutoff using chronological world event replay.

Self-Distilled Agentic RL extends on-policy self-distillation to multi-turn agents with token-level guidance and skill conditioning.

Dataset-agnostic audio framework converts text-based tool-calling benchmarks (Confetti, When2Call) to voice evaluations using TTS and noise without re-annotation.

VS Code adds Agents window for local AI models but requires internet + GitHub Copilot subscription, limiting offline/open-source utility.

SpeakerLLM integrates speaker understanding and verification reasoning for audio-first agents, conversational robots, and wearables via speaker-specialized audio-LLM.

Orchard open-source framework enables scalable agentic modeling with training infrastructure for autonomous LLM agents.

WARD defends web agents against prompt injection attacks via robust guard models and adversarial training.

AuthBench benchmark evaluates whether coding agents infer least-privilege file permissions for safe shell access.

CMBEvolve and CosmoEvolve demonstrate autonomous AI agents conducting scientific discovery in cosmology via code evolution and multi-agent workflows.

Websites can fingerprint LLM browser agents with 96% F1 accuracy via UI interaction traces, enabling targeted exploits.

π-Bench evaluates proactive personal assistant agents on identifying hidden user intents in long-horizon multi-turn workflows.

Automat: autoresearch framework using LLM coding agents to automatically design and optimize composition-based chemical descriptors for materials science.

Latent Space reports on rising usage of Codex and Claude for programmatic coding agents, noting continued industry trend.

I got tired of re-explaining the same codebase context to coding agents. Stuff like: “we tried moving auth into middleware, but backed it out because it broke OAuth callbacks,” or “that weird retry logic exists because Stripe webhooks arrive out of order.” So I built Almanac. It gives your coding agent a self-updating wiki for the codebase. It updates from your repo, and conversations you havewith Claude Code/Codex. The wiki lives locally in your repo as markdown. You can read it yourself, but the main consumer is the agent. It’s free and open source. Currently only MacOS (would add a wi...

Cohere showcases AI agents for financial services compliance, efficiency, and customer trust.

Notion’s new developer platform lets teams connect AI agents, external data sources, and custom code directly into their workspace as the company pushes deeper into agentic productivity software.

EVA-Bench provides end-to-end evaluation framework for voice agents, generating realistic bot-to-bot conversations and measuring voice-specific failure modes.

ScioMind integrates anchoring-based belief dynamics with LLM agents for cognitively-grounded social opinion simulation.

In today’s data-driven world, organizations increasingly rely on video to capture critical information, yet extracting meaningful, real-time insights from... In today’s data-driven world, organizations increasingly rely on video to capture critical information, yet extracting meaningful, real-time insights from massive amounts of footage remains a challenge. NVIDIA Metropolis Blueprint for video search and summarization (VSS) overcomes this hurdle by transforming millions of live video streams or hours of recorded video into instantly searchable… Source

Developer argues current AI agents require extensive human oversight and lack true autonomy despite productivity gains with Claude models.

LongMemEval-V2 benchmark evaluates whether agent memory systems enable agents to internalize environment-specific workflows and interface affordances in web tasks.

ToolCUA framework trains computer-use agents to optimally interleave GUI actions and tool API calls via trajectory-level supervision and synthetic data generation.

Text-tabular modeling predicts unfamiliar AI agent decisions in negotiation from limited interaction, tested on bargaining games.

Vapi says its enterprise business has grown 10-fold since early 2025 as companies shift customer support and sales calls to AI agents.

RobustToolBench benchmark exposes tool-use agent failures from deployment noise; domain-randomized RL improves robustness.

IPI-proxy toolkit enables red-teaming web-browsing AI agents against indirect prompt injection attacks embedded in whitelisted domain HTML.

GEAR enables fine-grained credit assignment in RL-trained LLM agents via adaptive-granularity advantage reweighting at token and segment levels.

The Karpathy coding skill is locked behind Pro. It doesn't use any Pro-only features, so I rewrote it for free plan chat workflows. Same philosophy, tuned for no terminal, no subagents, and a shorter context window where mistakes are expensive. Paste the whole thing into a Project's custom instructions or use it as a system prompt. It auto-triggers on any coding request. --- name: karpathy-coding description: Apply Karpathy-inspired coding discipline to any programming task. Use this skill whenever the user asks you to write, fix, refactor, extend, or review code — even casually...

Morning Everyone! Big one today (**104 changes!**): Claude Code just went async. The new `/goal` command lets you set a completion condition ("all tests pass and the PR is ready"), then Claude keeps grinding across turns until it's hit. The new `claude agents` view shows every session you've got running: working, blocked on you, or done. Translation: kick off a goal -> let claude cook -> come back later. First proper fire-and-forget loop CC has shipped. Pretty huge unlock if you've been juggling multiple sessions and losing track of which one needs you. Full notes: [https://www.luk...

OpenAI's Parameter Golf competition engaged 1,000+ researchers on AI-assisted ML workflows, coding agents, and model optimization under resource constraints.