2026-09-17
AI Agents
A serving-systems paper argues agents shouldn't guess how long a tool call will run — the running tool already knows, and reporting its progress cuts p90 time-to-first-token after a call by ~20% versus LRU eviction. Alongside it: a measurement of how much the coding-agent harness alone moves benchmark scores, a paper on running workflow portfolios under compute cost, and an open-source browser agent with an indexed action space.
Papers
2Ask the Tool, Don't Guess: Agent Tool Calls Hold Their Progress, and the Serving System Should Read It
The insight is sharp: agentic requests idle for wall-clock time during tool calls while their KV cache pins GPU memory, and serving systems decide eviction by guessing call duration from name, history, or declared timeouts. The paper shows no pre-call estimate can rank durations — but the running tool already holds the answer. A harness that surfaces per-call progress recovers a readable signal across four public corpora at no benchmark cost, cutting p90 TTFT after a tool call by ~20.7% against LRU, near oracle.
Designing Agentic AI Workflow Portfolios under Imperfect Selection and Compute Cost
Formalizes the run-many-workflows-then-select pattern as a portfolio problem: extra executions can catch answers the single best workflow misses, but they burn compute and introduce plausible distractors that degrade final selection. The authors summarize selector quality through an odds-lift index, derive bounds on the value of variety, and give LP relaxations plus a pricing-oracle method for large workflow classes. Evaluated on ABCD, Schema-Guided Dialogue, and HotpotQA; the abstract is truncated before the empirical results, so weigh the theory over the numbers.
News
2HarnessTax: How Much Does the Harness Matter for Coding Agents?
A direct measurement of how much scaffolding — not the model — drives coding-agent performance. This is the practical companion to yesterday's SWE-bench convergence audit, which showed scores swinging up to 29.8 points on scaffold choice alone. The HN thread is active (136 points, 55 comments) but the linked artifact is what matters: it quantifies the harness tax rather than treating scaffold as a footnote to the model name.
Jev Ultrafast: A browser agent with a dynamic, indexed action space
An open-source browser agent built around a dynamic, indexed action space rather than the usual flat DOM-to-action mapping. The framing suggests the index is rebuilt per page state, which could cut the token cost of describing available actions to the model on large pages. HN discussion is thin (65 points, 4 comments), so treat the performance claims as unverified until someone runs it against a browser-agent baseline.
30 more items the ranker flagged but didn't feature
Papers
- What Counts as Strategic Reasoning? A Systematic Mapping of Chess Research on Humans, Engines, and Language Modelsarxiv:arxiv-agents
- Affora: A Design System for Agent-Friendly Interfacesarxiv:arxiv-agents
- WFM: Wiki Foundation Model for Complex Agentic Reasoningarxiv:arxiv-agents
- Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommendersarxiv:arxiv-agents
- M-SQE: Multilingual Skill Quality Estimation for Enhancing Language Equality in Agentic Skill Usearxiv:arxiv-llm-systems
- An Empirical Evaluation of Cost-Efficient Large Language Models on Algorithmic Programming Tasksarxiv:arxiv-agents
- TuiML: Machine Learning for AI Agentsarxiv:arxiv-agents
- A Study on the Impact of Natural Language Differences in Prompts on Automatic Code Generation Using LLMsarxiv:arxiv-software-eng
- Flag Game: A Toy Model for Mechanistic Swarm Interpretabilityarxiv:arxiv-agents
- CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agentsarxiv:arxiv-agents
- RideWay: Benchmarking Efficient Task Completion for Tool-Using Language Agentsarxiv:arxiv-agents
- Faithful yet Collusive: Why Chain-of-Thought Monitoring Cannot Detect Collusion in LLM Pricing Agents under Oligopolistic Competitionarxiv:arxiv-agents
- From Component Snapshots to Lifecycle Traces: Agent-Based Software Composition Analysisarxiv:arxiv-software-eng
- A Study of the Reliability of Agentic AI-Generated Programsarxiv:arxiv-agents
- WetRobo: A Reproducible Robot Kit for Coding Agents in Biological Laboratoriesarxiv:arxiv-agents
- Who Audits Whom, on What Substrate, with What Evidence? An Independence-Graded Audit Protocol for Agentic AIarxiv:arxiv-agents
- Reasoning through Evolution: Automatic Meta-path Discovery for LLM-based Fake News Detectionarxiv:arxiv-agents
- Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environmentsarxiv:arxiv-agents
- AutoTuneBench: Trustworthy Measurement for Agent Auto-Tuning of LLM Serving Enginesarxiv:arxiv-agents
- CoRe-MARL: Cooperative Redistribution Under Unknown Dynamics Using Recurrent Multi-Agent Reinforcement Learningarxiv:arxiv-agents
- GraphEcho: Structural Redundancy and Evidence Provenance in LLM Graph Agentsarxiv:arxiv-llm-systems
- Recursive Reasoning or Statistical Extrapolation? In-Context Learning in Multi-Agent Interdependent Decision-Makingarxiv:arxiv-agents
- Collaborative Memory for Multi-Agent VLM Systemsarxiv:arxiv-agents
- Hypothesis-Driven Autonomous Materials Synthesis with Multimodal LLM Agentsarxiv:arxiv-agents
- Collective Loss of Control in LLM Agent Systems: An Epidemic Account of Mutation, Contagion, and Recoveryarxiv:arxiv-agents
- Social Laws for Multi-agent Coordination in Stochastic Environmentsarxiv:arxiv-agents
- DualSQL: Text-to-SQL with Multi-Agent Reinforcement Learningarxiv:arxiv-agents
- AeroWeaver: An Embodied-Agent Harness for Weaving Aerial Skills into Distributed, Adaptive Swarm Executionarxiv:arxiv-agents
- Market Signal Injection: Adversarial Context Manipulation of LLM Pricing Agentsarxiv:arxiv-agents
Blogs
- datasette 1.0a40rss:simonw