Agents Go Vertical: Domain-Specific Systems and the Race to Benchmark Reality

MedAgentBench: A Benchmark for LLM Agents on Real Clinical Tasks

A new benchmark tests agents on actual clinical workflows including diagnosis, treatment planning, and patient triage. This matters because healthcare is one of the highest-stakes verticals for agents, and domain-specific evals are the only honest way to know if a system is production-ready.

AgentReview: Automating Scientific Peer Review with Multi-Agent Systems

Researchers propose a multi-agent pipeline that simulates the peer review process, with agents handling different reviewer roles and meta-review synthesis. A useful proof-of-concept for high-context, judgment-heavy tasks that were previously considered too nuanced for automation.

ToolRAG: Retrieval-Augmented Tool Selection for Large-Scale Agent Tool Libraries

As tool libraries scale into the hundreds, naive tool selection breaks down fast. ToolRAG adds a retrieval layer that fetches the right tools at runtime, which is a practical fix for any builder dealing with bloated toolsets in production agents.

AutoCodeRover 2.0: Autonomous Software Engineering with Repo-Level Context

AutoCodeRover gets a major update with better repo-level context handling and improved patch generation across multi-file codebases. Software engineering agents are moving from toy examples to real PRs, and this is one of the more credible systems doing it.

AGENTMISTAKES: A Taxonomy of Failure Modes in Deployed LLM Agents

This paper catalogs real failure modes from deployed agents across planning, memory retrieval, tool execution, and user intent misalignment. A practical reference for anyone doing post-mortems on agent systems that aren't behaving as expected.