The Archive

Formal Skill introduces runtime-native abstractions for LLM agent skills with state management, policy enforcement, and completion discipline.

Developer adapts multi-agent orchestration patterns from Claude Sonnet to resource-constrained local setup using Qwen 3.6-35B.

DexHoldem benchmark evaluates embodied agents on real dexterous manipulation via Texas Hold'em poker with ShadowHand hardware.

EnvFactory scales tool-use agents via synthetic executable environments and robust RL, addressing data and execution bottlenecks.

SkillGenBench isolates skill generation as a benchmark task for LLM agents, measuring ability to create reusable executable skills from repositories.

Reversa framework converts legacy software into operational specs via multi-agent pipelines, enabling AI agents to safely modify existing systems.

Hugging Face reviving PapersWithCode with AI agents to auto-parse papers and generate leaderboards; covers Qwen, MTEB, ASR benchmarks.

Position paper argues agent generalization requires scaling environment rule-sets and interaction interfaces, not just trajectories within fixed benchmarks.

Empirical study of indirect prompt injection attacks enabling privacy leakage in black-box LLM chatbot agents with external tool access.

TaskGround framework for household agents to infer executable task structure from complete scene context and situated requests.

PPAI enables peer-to-peer collaboration between personalized LLM agents on edge devices via task delegation based on specialization.

DocOS framework enables GUI agents to proactively retrieve procedural knowledge from documents to handle long-tailed automation tasks.

I have been thinking about transactions in most agent frameworks. Consider an agent executing a sequence of five tool calls. If the third tool encounters an error, the resulting state is neither the user's intended outcome nor the system's state before execution began. Consequently, the agent has no systematic way to recover, and even a human operator must reconstruct what happened from incomplete evidence. This issue is not a problem with the tooling itself; it is a fundamental primitive missing from the stack. Databases have addressed this problem for 50 years, and distributed systems ha...

Reddit post claims multi-agent simulation with Claude, Gemini, Grok produced emergent behaviors; lacks peer review, reproducibility, or technical details.

I love Opus. First I tuned one page, got the PageSpeed result where I wanted it, and wrote the whole thing down in `ADR_pagespeed-l0-fixes-playbook.md`. Then I opened a fresh session, gave it the remaining 9 pages, and pointed it at the playbook. Opus created three subagents by itself, split the work between them, and about 15 minutes later they had touched 41 frontend files that powered those pages. Same result across the set. Basically perfect Lighthouse numbers again. Not gonna lie, this is the kind of workflow where I stop thinking “chatbot” and start thinking “tiny frontend team tha...

Anthropic recently published their [harness design for long-running apps](https://www.anthropic.com/engineering/harness-design-long-running-apps) — a multi-agent architecture inspired by GANs where a Generator builds code and an Evaluator critiques it in a loop. I built my own version using Kiro CLI and used it to generate a marketing website for my project [Mnemo](https://github.com/Mnemo-mcp/Mnemo) (persistent memory for AI coding agents). **The architecture:** Planner (runs once) → Generator ↔ Evaluator (12 iterations) Each agent is a separate CLI process with zero shared context. Th...

Mistral AI founder tells French Parliament that engineers now manage AI agents writing code instead of writing it themselves, marking a shift in developer workflows.

OpenAI announced yet another reorganization Friday, consolidating certain areas and making company president Greg Brockman the official lead of all things product. In a memo viewed by The Verge, Brockman wrote that since OpenAI's product strategy for this year is to go all-in on AI agents, the company is combining its products to "invest in a single agentic platform and to merge ChatGPT and Codex into one unified agentic experience for all." To do this, the company is making a suite of org chart changes, although it's still operating under some of the same ones from last month. That's when AG...

Framework combines generative AI agents and transformers for utility billing, carbon analytics, and load forecasting.

FORGE evolves natural-language memory for ReAct agents via population-based protocol without weight updates.

Argus: agentic system with Searcher-Navigator cooperation for evidence assembly in complex information-seeking tasks, addressing redundancy in parallel rollouts.

Benchmark of seven LLM tutoring agents on propositional logic reveals ceiling performance on correct solutions but systematic over-rejection of valid suboptimal answers.

paper.json: JSON companion format enabling LLM agents to reliably extract claims, scope, and reproducibility info from academic papers.

AI radio DJs demonstrated their volatile personalities. | Image: Cath Virginia / The Verge, Getty Images Andon Labs has been running a series of experiments in which AI agents run businesses without human intervention. Its latest is a quartet of radio stations run by some of the most popular AI models out there. "Thinking Frequencies" is run by Claude, "OpenAIR" by ChatGPT, "Backlink Broadcast" by Google's Gemini, and "Grok and Roll Radio," obviously enough, by Grok. They were each given a simple prompt: Develop your own radio personality and turn a profit…As far as you know, you will broadca...

Identifies premature exploitation in LLM agents; proposes Exploration Checkpoint Coverage metric and training methods for autonomous environment discovery.

ShopGym provides scalable benchmark environment for e-commerce web agents balancing realism and reproducibility across variable storefronts.

User reports local agent (Pi + Qwen 3.627B) executed rm -rf on full disk, demonstrating autonomous decision-making risk in unmonitored coding agents.

Reddit user argues Claude Code CLI is correctly positioned for personal coding tasks, not autonomous infrastructure agents.

OpenAI describes sandbox architecture for Codex on Windows, restricting file/network access for safe code execution in agents.

Coding agents reduce switching costs between programming languages and frameworks, enabling rapid rewrites (React Native, Bun); language/tool lock-in is eroding.