Agents Go Multimodal: Browser Control, Code Execution, and the Coordination Layer Nobody Agrees On

WebAgents Benchmark: How Well Do LLM Agents Actually Navigate Real Browsers?

A new benchmark directly tests agents on live browser tasks rather than synthetic environments, exposing a significant gap between lab performance and real-world reliability. If you're building anything browser-native, these failure mode breakdowns are worth studying before you ship.

CodeAct: Agents That Execute Code as Actions Outperform Tool-Calling Baselines

This paper argues that letting agents write and run code directly beats structured tool-calling in a majority of multi-step tasks. The implication is that sandboxed code execution should be a first-class primitive in your agent stack, not an afterthought.

Multi-Agent Coordination Without a Central Orchestrator: Decentralized Planning Results

Researchers tested fully decentralized agent coordination and found it competitive with centralized orchestration on several task types while being more fault-tolerant. The tradeoff is complexity in debugging, which remains a real operational problem.