AI Agents

This paper tackles agent memory for long-horizon tasks by moving beyond simple vector retrieval. The proposed HORMA architecture organizes experience into a file-system-like hierarchy, linking summaries to raw data. The key innovation is a second, lightweight agent trained with reinforcement learning to navigate this tree, selecting minimal but sufficient context. This "organize-then-retrieve" approach is shown to improve performance on benchmarks like ALFWorld under tight context budgets, directly addressing the cost and latency of ever-growing context windows.

Addressing the lack of training data for long-horizon coding tasks, this work introduces DeNovoSWE, a large-scale dataset of 4,818 instances for generating entire code repositories from documentation. The dataset was built automatically via a sandboxed agentic workflow using a "divide and conquer" and "critic-repair" philosophy, avoiding manual annotation. Fine-tuning Qwen3-30B-A3B on DeNovoSWE boosts its performance on the BeyondSWE-Doc2Repo benchmark from a baseline of 5.8% to 47.2%, demonstrating a massive leap in whole-repo generation capability.

This paper formalizes a common agent failure: "Silent Scope Omission," where a model applies a general rule but ignores nested exceptions. The authors propose a compiler-inspired solution by parsing rules into a Span-Grounded Deontic Tree (SG-DT). This intermediate representation explicitly models the control flow of rules and their defeaters, anchoring each logical branch to source text spans. Forcing models to generate this structured output is shown to significantly improve an agent's ability to correctly interpret complex policies from legal and corporate texts.

Naively increasing test-time compute (e.g., using a larger model or deeper chain-of-thought) for agent planners gives diminishing returns at high cost. This paper introduces DIRECT, a routing framework that decides when to allocate more compute based on multimodal scene context. The router dynamically chooses between different scaling axes—model size, thought depth, memory—on a per-prompt basis. On physical robot tasks, DIRECT matches a stronger model's success rate with up to 65% lower average latency, optimizing the cost-performance frontier.

For complex web search tasks, agents can get stuck pursuing a bad lead or waste time on random exploration. TreeSeeker introduces a disciplined "trial, error, and return" framework at inference time. It organizes the search as a tree, where each branch is a sub-goal attempt. Using textual signals for value, uncertainty, and risk (akin to UCB), the agent explicitly decides whether to exploit a promising branch, explore a new one, or prune a dead end and return to an earlier choice point.

Anthropic details how it enabled Claude to act as a research chemist, using a custom toolset of search, data analysis, and lab equipment APIs. This isn't just a demo; it's a deep dive into the agentic loop required for a complex scientific domain, covering planning, iterative refinement, and handling ambiguous results. The post offers a credible blueprint for building agents that need to chain multiple specialized tools to solve open-ended problems, moving beyond simple API calls and demonstrating a path for expert domains.