Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

agent harness

Think of an AI agent as a model plus all the supporting machinery around it: plans, tools, context management, verification, and security hooks. The model decides, but that machinery executes commands, checks completion, controls what the model sees, and guards against misuse.

Across the papers, harness design is a major lever. Planning guidance boosts accuracy, especially on hard tasks; verification cuts false successes but can block valid ones. Context management matters most with small memory budgets, and simple retrieval can beat multi-agent teams. Action languages let vision-language models control robots, while harness hooks open security holes. No component is universally best: its value shifts with model strength, task type, cost, and error risk.

Papers
176
Released code
9
First seen
June 2026
Latest
Nov 2026

156 papers in the last two months, against 19 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Terminal prompts make AI agents report success without doing the work

    Closing a command's input lets a software agent receive success status without the requested action being performed; reporting what the process waits on repairs this.

    Nov 2026
  2. 2Robots controlled by generalist AI outperform specialists without task-specific training

    Generalist AI models control manipulators more reliably than specialist policies, with 85.6% zero-shot success on RoboLab-120.

    Nov 2026
  3. 3Self-improving AI agents outperform fixed ones on open-ended tasks

    Recursive self-improvement helps frontier agents most when they build their own solutions and receive fine-grained feedback.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

agentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real42 togetherverificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.41 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.32 together

Who is working on it

NVIDIA 7University of Illinois Urbana-Champaign 6Zhejiang University 6Tsinghua University 6Alibaba Group 6Microsoft 6

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.