AI Agents

The PreAct framework accelerates agents on repeated tasks by compiling successful runs into compact state machines. Instead of invoking the full agent, PreAct replays the compiled program, achieving an 8.5-13x speedup by eliminating per-step LLM calls. This isn't blind replay; at each step, it verifies that the screen state matches the program's expectation and hands control back to the agent if anything deviates. A new program is only stored after being independently verified, preventing the accumulation of faulty automations.

This paper identifies a subtle but critical failure mode in multi-source agents: “cross-source conflation,” where an agent makes a factually correct claim but attributes it to the wrong source. The proposed solution, ProvenanceGuard, consumes structured agent traces (from the Model Context Protocol) to decompose answers into atomic claims and verify each one against its specifically attributed source. This source-aware verification allows for a block-or-repair loop, improving the factuality and trustworthiness of answers grounded in heterogeneous evidence.

A large-scale empirical study of 86,156 agent-authored test files on GitHub reveals that a staggering 80.2% contain weak or no explicit assertions (what the paper calls “oracle signals”). This means that while agents are successfully generating test files, these tests often execute code without actually verifying its behavior, creating a false sense of security. The study also found that while raw merge rates are lower for PRs with strong oracles, a regression analysis shows strong oracles significantly improve merge likelihood.

The new EComAgentBench benchmark is designed to evaluate shopping agents on more realistic long-horizon tasks. Its key feature is “distributed hidden intent,” where user requirements are scattered across an initial query, a tool-gated user profile, and scripted clarifications during the interaction. This forces agents to perform iterative information gathering. The results are sobering: even the strongest models tested only achieved 57.1% overall accuracy, demonstrating that current agents struggle to uncover and synthesize scattered requirements.

This paper provides a crucial empirical analysis of how agent tool routing degrades at scale. On a real-world enterprise system with 110 agents and 584 tools, routing F1 dropped by 16-23 percentage points as the catalog scaled. The authors decompose this degradation into a “retrieval gap” (failing to surface the right tool) and a “confusion gap” (failing to select the right tool from a good candidate set). An embedding-based shortlisting stage was shown to recover +10-11pp F1, confirming its value in production.

This work proposes a multi-agent framework for healthcare with two deterministic safety mechanisms to prevent common LLM failures. First, a neuro-symbolic state-tracker acts as a gate, ensuring a clinical data-gathering protocol (OLDCARTS) is complete before allowing a diagnostic step. Second, an uncertainty gate intercepts divergent outputs by calculating the semantic entropy across multiple independent samples. The combined system improved diagnostic precision by 11.3 percentage points over an unconstrained baseline, demonstrating a robust alternative to

This paper introduces a framework that automates prompt engineering for game-playing agents. It uses an LLM-driven evolutionary loop guided by environment rewards. The system features a “behavior analyzer” to attribute poor outcomes to specific parts of a prompt and a “mutator” to propose targeted revisions. This automated optimization process dramatically improves performance on the BALROG benchmark—on one task, it raised the success rate from 0% to 72.5% using the same underlying LLM, purely through prompt refinement.

This paper introduces Tensor-Coord, a highly formal framework for multi-agent coordination using multilinear algebra. It represents the joint plan of N agents as a tensor and uses decomposition techniques (CP, Tucker) to identify a plan's “coordination complexity” and localize conflicts without domain-specific rules. The mathematical outputs are then converted back into natural language constraints to guide an LLM in an iterative replanning loop, successfully generating conflict-free plans for multi-robot tasks.

This paper argues that standard predictive accuracy metrics are insufficient for validating agentic systems. It proposes a formal validation framework based on Partially Observable Markov Decision Processes (POMDPs) that decomposes agent behavior into distinct components: belief state formation, forecasting, and policy selection. This allows each part of the agent's decision-making process to be validated independently, providing a structured and rigorous methodology for assessing and managing the model risk of autonomous systems.

To handle complex tasks that require multiple tools, this paper introduces SkillWeaver, a decompose-retrieve-compose framework. Its key innovation is “Iterative Skill-Aware Decomposition” (SAD), a retrieval-augmented feedback loop. An LLM first proposes a task decomposition, which is then refined based on which skills are actually available in a tool library. This single refinement iteration improved decomposition accuracy by over 32%, ensuring the agent produces plans that are grounded in its real capabilities.

Georgi Gerganov, creator of ggml and llama.cpp, shares his personal setup for using local LLMs as a coding assistant. He uses a Qwen3.6-27B model on either an M2 Ultra or an RTX 5090, driven by a stripped-down, offline agent (pi -nc --offline) and a custom system prompt to align the output with his style. It's a strong, practical endorsement from a key practitioner for using local models to handle mundane but helpful daily maintenance tasks in a large open-source project.

The latest alpha release of Datasette (1.0a34) adds UI controls for inserting, editing, and deleting database rows. The most interesting part is the motivation: the author added SQL write capabilities to Datasette Agent and realized how absurd it was that a chat agent could modify the database but the core GUI could not. It's a great example of how building agent capabilities for a product can directly surface and inspire long-overdue improvements in the conventional user interface.