Open Weights, Bot APIs, and the Shift to Agent Liability

Research

IoA: An Interoperability Protocol for Heterogeneous Agent Teams

The 'Internet of Agents' is a robust protocol for orchestrating heterogeneous agents over complex networks. For builders, this means finally having a standardized way to get specialized agents to cooperate on complex tasks without tight coupling.

Verifying LLM Agent Policies Against Norms and Values

We write rules for agents all day, but few know if they actually work. This paper provides a grounded, graph-theoretic way to verify those 'house rules' (policies) against a specific ontology, which is crucial for building safe, constrained agents.

Analysis

The New Reality of AI Coding Coworkers

Mollick argues that AI coding tools are now good enough to fundamentally change the dev workflow. If you aren't treating AI as a 10x coworker rather than a fancy autocomplete, you're leaving leverage on the table.

News

Kimi K2: Open-Source MoE Model Outperforms GPT-4.1

Kimi just dropped an open-source MoE model with a massive 1T total param count that punches way above its weight class. This is serious competition for closed-source giants and validates the open-weight approach for high-performance agents.

Tools

Firecrawl Launches Agent-First API Endpoints

Firecrawl just dropped a specialized endpoint designed specifically to feed Agents, not just humans. It prioritizes structured, machine-readable data over pretty HTML, which solves a massive friction point for autonomous browsing loops.