Agents Get New Legs: Browser Control, Multi-Agent Benchmarks, and the Tool-Calling Wars

News

OpenAI Expands Operator Agent Capabilities with Improved Web Task Completion

Operator is getting better at multi-step web tasks, which puts real pressure on open-source browser automation alternatives. If this matures, it shifts the question from 'can agents browse the web' to 'which agent does it with the fewest failures.'

Anthropic Tool Use Hits General Availability with Expanded Function Calling Support

Claude's tool use is now production-ready with better handling of parallel tool calls and improved JSON schema validation. For anyone building agents on top of Claude, this removes a lot of the defensive code you had to write around unreliable function calling.

Research

New Benchmark Reveals Multi-Agent Systems Still Struggle with Long-Horizon Task Planning

Researchers dropped a new evaluation suite showing most multi-agent frameworks fall apart when tasks require more than 10 sequential steps. This is the kind of honest signal builders need before committing to a framework for production workflows.

Tools

AutoGen 0.4 Ships with Improved Actor-Based Agent Runtime and Distributed Execution

Microsoft's AutoGen overhaul moves to an actor model under the hood, making it easier to run agents across distributed processes without managing state manually. This is a meaningful architectural shift for teams building anything beyond toy demos.

Docker Publishes Security Patterns for Sandboxing AI Agents in Production

Running agents that execute code or browse the web without proper sandboxing is a security liability most teams are ignoring right now. Docker's guidance here is practical and fills a gap that most agent framework docs completely skip.

Analysis

HuggingFace Analysis: Which Open-Source LLMs Actually Work as Agent Backbones

HuggingFace ran systematic evals on open-weight models for tool calling and multi-step reasoning, and the results are more nuanced than most vendor benchmarks admit. Useful if you're deciding whether to pay for frontier APIs or self-host for agentic use cases.