Paper Overview
Research area: cs.AI / long-horizon agents / harness evolution Authors: Lingzhi Yuan*, Yiguang Cai*, Weize Liu, Siqi Yang, Feijiang Han, Yongkyun Lee, Udari Madhushani Sehwag, Furong Huang (University of Maryland, University of Washington, UIUC, Scale AI) arXiv: 2611.05139 Project page: https://meta-square.github.io/
Background: The Horse and the Harness
The performance of a long-horizon agent depends not only on the model but on the harness wrapped around it: how prompts are written, how context is trimmed, which tools are registered, where hooks fire, how completion is checked, and how memory is organized. The trouble with harnesses is that you only learn where they fit poorly *after* the horse starts running—missing tools, missing constraints, and needed routine checks only reveal themselves mid-run.
Existing harness-evolution approaches split into two camps:
- Offline methods distill all training experience into one shared harness h*. This amortizes experience but can overfit training tasks, and once deployed, the harness is frozen—unable to react to evidence revealed during a new run.
- Online methods adapt before, during, or after a run. Within-run adaptation is the most responsive, but the evolver must reason about every intervention from scratch (burning tokens, latency, and money at every checkpoint), or rely on verifier feedback unavailable at test time.
- Condition A: task agent alone, no evolver (baseline).
- Condition B: evolver guided by the current Meta-Skill.
- Condition C: evolver given privileged verifier information available only during training.
- B fails, A passes: the evolver hurt—"avoid" evidence.
- B passes, A fails: worthwhile interventions—"act" evidence.
- C passes, B fails: missed recovery strategies—additional act evidence.
- C fails, B passes: inherently toxic interventions—blacklisted.
- A passes, B passes: the agent handles it alone—"hold" evidence: the correct action is to do nothing.
- Performance (GPT-5.6-Terra): Meta² reaches 81.00% on Terminal-Bench (seed: 66.33%, +14.67; best baseline Meta-Harness: 73.33%) and 44.44% on DeepSWE (seed: 34.67%, +9.77). On DeepSWE, ReasoningBank, ACE, and Continual Harness fall below the seed; only Meta² improves substantially, attributed to condition C providing "could-have-succeeded" contrasts on hard tasks.
- Different backbone: DeepSeek-V4-Flash-0731 achieves 77.67% on Terminal-Bench (+7.67 over seed). Each backbone learns its own Meta-Skill, so this tests mechanism robustness rather than zero-shot skill transfer.
- Cost: Terminal-Bench total $174.91 ($110.19 offline + $64.72 online) vs. AHE $248.35 and Meta-Harness $198.85. DeepSWE total $367.86 vs. AHE $862.89 and Meta-Harness $870.75. The offline cost is a one-time investment amortized across all deployments.
- Checkpoint frequency: L=5 is the sweet spot (81.00%)—too frequent interrupts the agent; too sparse misses intervention windows.
- Ablation: online evolution alone helps; adding contrastive Meta-Skill learning helps further—components are complementary.
- The Meta-Skill is backbone-specific: switching models requires re-learning; experiments test cross-backbone robustness, not zero-cost transfer.
- The one-time offline cost ($110–$238) pays off only for streaming workloads, not one-off tasks.
- The checkpoint interval L=5 may not generalize to other task lengths or environments.
- Meta-Skill quality is bounded by training-task coverage; unseen failure modes cannot be distilled.
- Meta² optimizes evolution decisions, not model capability—the model's ceiling remains.
- Condition C's privileged information requires strong information-isolation guarantees and trajectory auditing, which grows harder in complex tool environments.
- Yuan L., Cai Y., Liu W., Yang S., Han F., Lee Y., Sehwag U. M., Huang F. "Meta²: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents." arXiv:2611.05139, 2026. https://arxiv.org/abs/2611.05139
- Project page: https://meta-square.github.io/
Meta² asks: can completed, verified training experience guide online local edits without importing task-specific modifications into unrelated tasks?
The Meta-Skill: Storing the Trial-and-Error
Meta²'s answer is a middle path: instead of optimizing a shared harness, precompute a reusable Meta-Skill (𝓜)—knowledge about *when* to evolve, *how* to evolve, when not to touch anything, and which interventions to avoid. If the harness is a recipe, the Meta-Skill is the culinary know-how of when and how to modify recipes and what pitfalls to avoid.
Online component: checkpoint-based harness evolution
The task agent starts from a minimal seed harness h⁽⁰⁾. At each checkpoint, the evolver receives new trajectory evidence Δτ⁽ᵏ⁾ plus read-only workspace observations ω⁽ᵏ⁾, with accumulated information:
where σ⁽ᵏ⁻¹⁾ records the fate of the previous edit (loaded, kept, or rolled back). The evolver makes a binary decision:
Edits flow through a bounded channel: a system notification, an instruction update, a hook injection, or a tool registration—never a full harness rewrite. At the terminal checkpoint, EVOLVE becomes a bounded correction that names a specific contradiction and scopes the repair to that contradiction. Every edit applies only to the remainder of the current run and is discarded when the run ends.
Offline component: learning the Meta-Skill from three contrasting conditions
Rather than collecting successful trajectories (since a run's outcome is a joint evaluation of all checkpoint decisions), Meta² runs each training task under three conditions:
Five contrast directions yield evidence types:
A strict red line prevents condition-C prompts from leaking privileged verifier information to the task agent; trajectories are audited and violations removed. Each batch then passes through three LLM stages—REFLECT (write candidate entries against a schema), AGGREGATE (merge cross-task patterns, strip task-specific details), and UPDATE (add/merge/modify/delete). The final Meta-Skill is frozen—read-only from deployment onward.
Deployment: freeze globally, adapt locally
At test time, checkpoint decisions condition on:
No online decisions, task-specific edits, or run outcomes are written back to 𝓜. Knowledge is reused across tasks; edits dissolve with each run, preventing one task's corrections from polluting unrelated tasks.
Results
Benchmarks: Terminal-Bench 2.1 and DeepSWE 1.1 (random 1:2 train/test splits), mean@5 / mean@3. Baselines include hand-designed harnesses (NexAU, Codex CLI, mini-swe-agent, Pi agent, DeepSeek Harness), self-evolution methods (ACE, Training-free GRPO, ReasoningBank), and harness-evolution methods (AHE, Meta-Harness, Continual Harness), all from the same NexAU seed.
What Meta-Skill Entries Look Like
Each entry has: type (act/hold/avoid), timing (begin/middle/end), trigger conditions described via observable trajectory evidence, scoped guidance, rationale, and reference task IDs. Example: an act/end entry flags "the delivered artifact contradicts its own trajectory"—when an agent claims completion but its deliverable contradicts the execution trace (e.g., the last pytest run predates a code change), send it back for a repair scoped strictly to that contradiction, since budget is scarce at this stage.
Limitations
Takeaway
Meta² turns "how to modify a harness" into a reusable asset: offline, a Meta-Skill is distilled from three-condition contrastive rollouts; online, knowledge stays frozen globally while edits are applied locally and discarded after each run. The results frame harness evolution as an amortizable infrastructure investment rather than a from-scratch craft at every deployment—and a step toward recursive self-improvement with a built-in leash: every improvement must prove itself within the current task's sandbox.
---