Your policy is secretly a harness; self-evolving harnesses (Trinity, EMMA, HaMo) and their limits

Sources (HuggingFace 2026-10-28 + X feed): Your Policy Is Secretly a Harness, 2610.37255 · The Trinity, 2610.36077 (HF #1, 535 upvotes) · EMMA, 2610.35966 (Amazon, via @omarsar0 and DAIR) · HaMo, 2610.34724 · Are self-evolving agents continual learners?, 2610.25858 (via @xiaoyin_chen66) · h1, 2610.35962 · TTME, 2610.32046.

TL;DR. The harness (the prompts, tools and control logic around a model) is now treated as a trainable object, and today's papers say how to measure it and where self-editing breaks.

• Your Policy Is Secretly a Harness. Under KL-regularised control, a trained policy's log-odds recover how strongly its harness favoured each action, so the model carries its harness's preferences. Deployed return = the harness's ceiling (best return it offers) minus the policy's mismatch (return lost to its own choices). Training changes only mismatch; changing the harness moves both. So "compare harnesses with a fixed model" confounds the two. On MiniWoB++, merely reversing the order of observation fields (no optimal decision changes) drops a trained model from 85.0% to 19.2%. Harnesses a fixed-model comparison rejected but that raise the ceiling beat the original after retraining; ceiling minus mismatch (with a stronger model standing in for the optimum) reproduces the Terminal-Bench Lite ranking where neither term alone does.
• The Trinity. A Creator / Explorer / Evolver loop reorganises tasks by shared capability rather than surface category, builds simplified verifiable simulated worlds per capability, and evolves a tree of reusable harness modules. A capability-aware harness router picks a module configuration per task. +34 points mean across 28 training benchmarks and +26 across 18 unseen ones.
• EMMA (Amazon). Treats a multi-agent system's controller routing memory, specialist prompts and per-step skills as episodic memory. After a failure, a reflector decides which of the three owns the fix and proposes one scoped, bounded edit. Beats ACE, GEPA, EvoMAS and ADAS by up to 32.5% on KramaBench, AppWorld and BBEH with 17-38% fewer training executions.
• HaMo. Fast non-parametric "experience cards" for immediate reuse plus slow parametric training that teaches the model to execute retrieved strategies. Beats harness-only and model-only adaptation under domain shift.
• Counter-evidence: self-evolving agents are not continual learners. ReasoningBank, GEPA and Meta-Harness, run on four drifting task streams (BIRD, BFCL, ScienceWorld), improve aggregate scores but trade stability for plasticity: global edits to a shared prompt overwrite earlier knowledge as the stream shifts.
• h1 (bitter lesson for theorem proving). A plain loop with an efficient Lean 4 interface matches or beats specialised workflows (Goedel-Architect, AlphaProof Nexus) at matched models and settles eight open FormalConjectures problems. The interface, not the workflow, sets cost and solve rate.
• TTME. Evolving an external memory at test time beats test-time training on five math discovery tasks at 82.8-84.6% fewer FLOPs.

flowchart LR
  H["Harness<br/><small>prompts, tools, order</small>"] --> C["Ceiling<br/><small>best possible return</small>"]
  H --> P["Trained policy<br/><small>inherits harness prefs</small>"]
  P --> M["Mismatch<br/><small>return lost to choices</small>"]
  C --> R["Deployed return<br/><small>ceiling minus mismatch</small>"]
  M --> R
  P -.->|reorder fields| X["85.0% to 19.2%<br/><small>brittle preference</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class H input
  class P core
  class C,M loop
  class R exit
  class X err
  linkStyle 5 stroke:#e03131,stroke-width:2px

Relation to prior wiki

• Explains a pattern in agent harness engineering: the 10-24 harness-cost study found 3x cost spread across harnesses at equal pass rate, and the 10-28 Trial by Trace benchmark scored harnesses rather than models. Ceiling vs mismatch says why fixed-model harness comparisons mislead.
• N-of-a-kind on routers inside harnesses: Trinity's capability-aware harness router, EMMA's controller routing memory, and Copilot Fusion's cross-lab model router all put a routing decision inside the harness. See LLM routing.
• Contradiction, unresolved: Trinity and EMMA report large gains from self-editing harnesses; the continual-learning paper finds the same family forgets under drift. EMMA's scoped single edits are the most plausible fix, but it was not tested on drifting streams. See self-evolving agents.
• Bitter-lesson echo: h1 joins Axiom's CFSG formalisation (10-28), which credited harness design over model size.