Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

harness evolution

When you give an AI agent a complex job, you first build its "harness", the prompts, tools, and rules it follows. Evolution means automatically tweaking that harness, trying different versions, and keeping the ones that work better, all while leaving the AI's internal model unchanged.

Across the papers, the common thread is making harness evolution more efficient and reliable. They tackle two big problems: wasting evaluation effort on easy tasks, and overfitting to specific mistakes. Some methods sample tasks smartly to cut costs by up to 80%, others focus on failure patterns that repeat across many tasks, and some adapt the harness itself as the AI learns. The result is that evolved harnesses get better scores, transfer well to new tasks or different AI models, and require less manual engineering, turning harness tuning from a one-time human job into a continuous, automated process.

Papers
66
Released code
5
First seen
June 2026
Latest
Nov 2026

52 papers in the last two months, against 10 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Models can build their own agent software, but improvements generalize unreliably

    LLMs can build capable task harnesses, but two rounds of self-revision do not reliably improve held-out performance and may trade performance for lower inference cost.

    Oct 2026
  2. 2Robot agents approach teacher performance using one fifth the accelerator time

    Harness evolution for a large teacher can be distilled into a compact student that achieves comparable success at one-fifth the compute.

    Nov 2026
  3. 3AI agents complete roughly four times as many tasks with helpers

    Reusable primitives with embedded LLMs, composed per task by HEELO, solve real-world tasks that frontier models fail to complete.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.31 togetherself-evolving agentAn agent that improves by reusing its own experience, memory, or self-generated feedback, with success depending on verification and design.18 togetherfailure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.17 together

Who is working on it

Georgia Institute of Technology 3Shanghai Jiao Tong University 3Amazon 3NVIDIA 3Zhejiang University 3Microsoft Research Asia 3

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.