AI Agents

This paper introduces the "Queen-Bee" architecture, a multi-agent pattern designed for governed enterprise environments. The system separates planning from execution: a central "Queen" agent handles task planning and compiles a structured plan called a "BeeSpec." This spec is then executed by specialized "Bee" agents that operate with constrained, tenant-scoped tool access. This design enforces policy and operational boundaries by construction. On 59 enterprise tasks, the system achieved a 96.4% success rate with zero governance failures, outperforming a permissive single-agent baseline.

The DuMate-DeepResearch framework presents a multi-agent architecture for complex, open-ended research tasks. It introduces three notable mechanisms to improve long-horizon planning. First, it uses a dynamic graph for coarse-to-fine planning that allows for reflection and re-planning. Second, it employs a recursive design, delegating complex sub-tasks to a dedicated "Search Agent" with its own internal planning loop. Third, it generates task-specific quality rubrics at runtime to ground its reasoning, making the entire process auditable by design.

This paper applies concepts from Byzantine Fault Tolerance (BFT) to multi-agent LLM collaboration. Traditional BFT requires byte-identical messages, which LLMs don't produce. The proposed protocol, Hierarchical Certified Semantic Commitment (H-CSC), instead uses embedding similarity over agent proposals to reach a decision. It produces one of three typed outcomes: a full semantic_commit if agents agree on the rationale, a verdict_commit if they only agree on the outcome, or a typed abort. It's a formal approach to achieving consensus among unreliable semantic agents.

MemDreamer is a framework that tackles the token limits of long video understanding by decoupling perception from reasoning. It first processes a video to build a hierarchical graph memory, capturing spatiotemporal and causal relationships. A separate reasoning agent then interacts with this memory using tools in a classic Observation-Reason-Action loop, navigating the graph to answer questions. This agentic retrieval approach achieves SOTA results on four benchmarks while using only 2% of the context window required by a full-ingestion model.

The NTILC paper offers a solution to the scaling problem of tool use: as your tool library grows, stuffing all function signatures into the context window becomes untenable. NTILC replaces this in-context approach with retrieval. It learns a shared embedding space for user intents and tool specifications, allowing the agent to perform a retrieval step to find the most relevant tool. The LLM is then prompted with only that single tool's schema to generate arguments, reducing context use by over 95%.

A new proposal, the Universal Memory Protocol (UMP), aims to establish a shared, open standard for agent memory. The goal is to solve a fundamental interoperability problem: how can different agents, potentially built with different frameworks, share and understand each other's memories? By defining a common format, UMP could enable more seamless collaboration and state transfer between agents. The specification is currently in an early proposal stage, inviting community feedback on a critical piece of agent infrastructure that could benefit the whole ecosystem.

A new paper (arXiv:2602.11988) explores whether providing an agents.md file in a codebase can improve the performance of coding agents. The idea is simple: just as a README.md guides a human developer, an agents.md provides high-level context, architectural overviews, and instructions specifically formatted for an LLM agent. The paper presents this as a technique for improving agent planning and task execution, essentially a form of structured, in-repo guidance that's more direct than relying on source code and comments alone.

Lathe is a new open-source tool that flips the script on LLM-driven development. Instead of having an agent do the work for you, Lathe uses an agent to generate a detailed, hands-on tutorial so you can learn a new technical domain yourself. You prompt it with a topic (e.g., "build a 3D slicer in Erlang"), and it produces a local webapp with a full tutorial, complete with code, exercises, and source links. The emphasis is on active learning by typing the code by hand.

Simon Willison has released datasette-agent-edit, a new plugin that provides a reusable toolkit for agentic text editing. Inspired by the design of Claude's text editor tool, it implements core functions like view with line numbers, str_replace, and insert at a specific line. By factoring these common patterns into a base plugin, Willison aims to create a standard foundation for more advanced agent tools that need to edit text, from Markdown documents to SQL queries, without each having to reinvent the same low-level primitives.