Trends

Agents improve themselves through skills, harnesses, and memory

Agent training and memory · Since Jul 31

Organizations

MeituanAlibabaGoogleMicrosoftNVIDIAShanghai AI Lab

Updates

  1. Oct 2

    ActiveSaddler co-evolves the training curriculum with the harness, PoS keeps explicit belief states for long tasks, and Vast-10M pitches 10M-token context as in-context continual learning.

    Microsoft's ActiveSaddlerTsinghua and Alibaba's PoS belief statesVoltropy's Vast-10M with 10M-token context

  2. Sep 25

    Growing Harness evolves an executable harness from failures and cuts deployed LLM calls 76-92%, and Pistis tunes its inference harness automatically without touching weights.

    Growing Harness evolves agent scaffolds from failuresPistis pairs IDRL with auto-harnessing

  3. Sep 18

    NVIDIA's SoL-Pi used automated research loops to find harness mechanisms that match Pi on EdgeBench while cutting token traffic 44.7-49.0% and API cost by about a third, and a 176-setting study showed which context and planning choices pay off on SWE-Bench Verified and Terminal-Bench 2.1.

    NVIDIA's SoL-Pi automates harness researchEmpirical study of coding-agent harness design

  4. Aug 28

    Google's WikiSkill evolved skills that transfer across model families so smaller models beat larger ones, TTPO lifted Qwen3-1.7B from 38.0% to 45.2% on competition math without labels, and AHEAD's step-aware hints added 13.3 points on ALFWorld and 11.0 on WebShop.

    Google's WikiSkill co-evolves skills and knowledgeZhejiang University's label-free TTPOAHEAD adds step-aware hints to agentic RL

  5. Aug 14

    Meituan's AutoDesign has a code agent recursively rewrite its own harness, lifting PosterBench from 54.99 to 67.39 and beating Claude Design by 7.45 points with a frozen model, and SKILLER's generated skills add up to 20.4 points to small open models.

    Meituan's AutoDesign rewrites its own harnessOpenDataLab's SKILLER writes skills for small models

  6. Jul 31New

    SkillRise has one policy curate a skill document across tasks for 2.3-8.5 Pass@1 gains, Meituan's CAST turns solver value changes into dense turn-level rewards, and DecoEvo co-evolves solver and rubric skills to beat prior text-space optimizers by 2.8-5.0%.

    SkillRise curates evolving skills across agent tasksMeituan's CAST turns solver values into rewardsDecoEvo co-evolves solver and rubric skills

Predictions

  1. Jan 2027

    Skill curation and process-level credit assignment become default components of agent training stacks, lowering the environment and rollout cost of building reliable agents. Jul 31

  2. Feb 2027

    Harness and skill search become a standard way to convert evaluation compute into durable capability from frozen models, reducing reliance on retraining for incremental gains. Aug 14

  3. Feb 2027

    Skill and memory stores plus label-free test-time training become standard layers of the agent stack, decoupling incremental capability gains from expensive pretraining and letting smaller models punch above their weight. Aug 28

  4. Mar 2027

    Harness and inference-efficiency engineering, not only base-model quality, becomes a primary lever on agent cost-performance, and recursive self-improvement at the harness layer moves toward production. Sep 18

  5. Mar 2027

    Harness and inference-efficiency engineering become a primary cost-performance lever, letting smaller deployment models match larger ones. Sep 25

  6. Apr 2027

    Belief-state and memory management plus curriculum-driven harness search become the next cost-performance levers, narrowing the gap between small deployment models and larger ones. Oct 2