A single AI agent requires a harness—a control layer managing context, tools, execution, and verification. However, an enterprise fleet of agents needs an operating system to manage contention and resource allocation across multiple agents competing for shared resources, fundamentally different from scaling up a single harness.
Artificial Analysis's Intelligence Index, a widely-used AI model leaderboard, compresses diverse evaluation choices into a single intelligence score that obscures methodological trade-offs. The index weights agentic workloads heavily (34%), relies on saturated benchmarks like GPQA that cannot distinguish frontier models, lacks genuine coding benchmarks despite a 24% coding category, and features half its components using similar agent-execution patterns that may over-represent certain capabilities.
Cayu is an open-source Python framework for building domain-specific long-horizon agents with durable state management. It enables agents to process complex, multi-step tasks across long sessions, pause for human input, and recover from recorded evidence without losing progress. The framework handles session recovery, permissions, budgets, and observability for enterprise workflows.
Weave is experimenting with using AI agents on a prediction market exchange to price the impact and utility of code pull requests, leveraging a central limit order book where agents can buy and sell based on their assessments of a PR's value relative to engineering time investment.
Pixel Agents is a pixel-art mission control interface for Claude Code agents that allows real Claude sessions across app and terminal, displays project status on a pannable map, manages shared task lists, tracks exact costs, and was built with agents drawing their own pixel-art sprites.
Researchers replaced human subjects with LLM agents in double auction market experiments and found that LLM-populated markets converge more slowly to equilibrium than human markets, resulting in less efficient resource allocation. Analysis revealed significant heterogeneity in trading behavior across model families and roles, with agents shifting from strategic reasoning to urgency when deciding to trade.
Over a decade ago, developers emerged as essential power brokers in technology. Now, AI agents are fundamentally reshaping that landscape by autonomously writing code, selecting languages and frameworks, and making architectural decisions—shifting the role from human developers as kingmakers to machines as decision-makers.
Inwom is a security tool designed to detect coordinated behavior among AI agents and machines by analyzing collective activity patterns rather than individual events. It reconstructs how agents organize, share capabilities, and act as coordinated systems using observable evidence from trusted infrastructure, distinguishing legitimate automation from potential attacks.
Scaling Trust is a £50 million UK R&D programme launching the Arena, a physical experimental testbed in the UK going live in early 2027 where AI agents operate autonomous organisations in a simulated economy, competing and interacting under adversarial conditions with a multi-million pound prize pool for top performers and red teamers.
Chert FaceTime Opensource is an open-source CLI and SDK that enables developers to deploy AI agents on live FaceTime calls using WebRTC and LiveKit. Users retain control over their agent's code, model, and voice while running it locally or on their own infrastructure, with audio and video bridged between FaceTime and a LiveKit room.
Mayfly Chat is a lightweight tool that creates transient chat channels for AI agents to communicate directly with each other, eliminating the need for manual intervention. The creator describes various use cases like agents coordinating configuration tasks across machines and working together on distributed jobs, while emphasizing that end-to-end encryption via client-side scripts was implemented as a security requirement for public deployment.
Chat-Man is a free WhatsApp MCP server that enables AI agents to read, search, extract, and send WhatsApp messages, along with a web UI for managing groups and contacts. It allows users to integrate WhatsApp access across multiple projects without separate library integrations, supporting features like message summarization, CRM record creation, and bulk messaging.
Delta is a multiplayer environment for collaborative coding with AI agents that unifies conversations, code changes, and version control in shareable threads. It enables real-time teamwork with anchored comments and decisions, backed by DeltaDB, a version control system that tracks fine-grained changes rather than snapshots.
Anthropic launched Delta, a public beta multiplayer coding environment designed to replace pull requests by enabling real-time collaboration between developers and AI agents. Delta uses DeltaDB, which extends Git with delta-based versioning to preserve the evolution of code changes alongside agent and human context, allowing teammates to join threads and continue work without traditional commit-based workflows.
X posts discuss DeFi developments on Arc blockchain, including multiple launchpad and DEX projects, Solana's Transaction V1 launch enabling atomic multi-step trades, and speculation about AGI emerging from distributed AI agent networks rather than a single model.
Two cryptocurrency projects are gaining attention on X: CyberThrone NFTs are expanding to multiple blockchains including Robinhood Chain through Unvault's LayerZero technology, allowing collectors to mint and transfer assets across chains. ANT is developing a multi-agent AI system inspired by ant colonies that coordinates specialized agents to complete complex tasks, with plans to launch its $ANT token on Robinhood Chain.
BrowserSkill from Tencent is a browser automation tool that allows AI agents like Claude Code and Cursor to control your browser without interrupting your work. It enables agents to reuse existing login sessions, run tasks in a separate agent window, and includes human-in-the-loop support for tasks requiring manual intervention like CAPTCHAs.
Solana launches Transaction V1 on mainnet enabling atomic multi-step DeFi trades with ZK-proofs. Discussion highlights blockchain infrastructure positioning for AI agents and autonomous commerce, with Sui emphasizing programmable assets and atomic execution capabilities for machine-driven financial transactions.
Six months after publishing "Large Language Models for Mortals," the author recompiled the book's 264 executable Python code cells and found minimal breakage: only three retired models required updates (Claude Opus 4, Claude 3 Haiku, Gemini 3 Pro), while core APIs like chat completions and structured outputs remained stable. The author notes what held up well (fine-tuning predictions, RAG patterns, TDD for agents), what needs updating (AWS Bedrock's simplification, open-source model endpoints), and upcoming challenges like temperature=0 deprecation in newer thinking models.
This article explains how to build effective evaluations for AI agents, covering task design, grading methods, and eval harnesses. It highlights why agent evals differ from single-turn LLM evaluations due to multi-step reasoning and compounding errors, and recommends separating failures into reasoning, action, and execution layers. The piece provides practical guidance on sourcing tasks, writing clear success criteria, and tracking performance changes over time.