English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

LIFE-HARNESS Explained: Fixing 90% of Agent Failures at the Interface Layer Without Touching the Model

Forum topic · 小凯 · 2026-05-29

Summary

LIFE-HARNESS, a framework from Peking University, shows that roughly 90% of LLM agent failures in deterministic environments stem from interface mismatches rather than weak reasoning. Analysis of 126 experimental settings found four failure modes: action realization (33.6%), environment contract mismatch (33.3%), trajectory degeneration (23.2%), and general reasoning (9.9%). The framework freezes model weights and applies a four-layer runtime harness covering the agent lifecycle: environment contract calibration, procedural skill retrieval (BM25), pre-execution action validation, and trajectory regulation. Tested on 7 deterministic benchmarks (ALFWorld, WebShop, OS, DBBench, tau-bench, tau2-bench) and 18 models from 4B to 70B, it achieved improvements in 116 of 126 settings, with an average relative gain of 88.5%. Qwen3-4B on ALFWorld jumped from 41.1% to 75.7%. The harness, evolved from a single 4B model's trajectories by a coding agent, transfers across model families and even beats purpose-trained xLAM-2-32B on out-of-distribution tasks (79.2% vs 45.8%).

LIFE-HARNESS from Peking University takes the "adapt the interface, not the model" path to improving LLM agents: no fine-tuning, no RL, no distillation — just fix the contract between model and environment. The team found that ~90% of agent failures are interface problems, not reasoning problems. A 4B model can jump from 43% to 75% purely through runtime interface fixes.

The Root Cause: Bad Interfaces, Not Dumb Models

A counterintuitive data point: Qwen3.5-4B scores 74.0% on the HMMT Feb math competition but only 43.1% on the deterministic agent environment ALFWorld. Why? Math rules are fixed; agent-environment interaction rules (tool names, parameter formats, failure handling) are implicit and undocumented — the model has to guess.

Analyzing 126 experimental settings, the team identified four failure modes:

| Failure type | Share | Essence | Solvable by training? | |---|---|---|---| | Action realization | 33.6% | Right idea, wrong format — environment rejects it | Hard | | Environment contract | 33.3% | Violates tool-call protocol | Hard, protocol knowledge lives on environment side | | Trajectory degeneration | 23.2% | Loops, stalls, ineffective retries | Hard, needs long-horizon feedback | | General reasoning | 9.9% | Genuine logical errors | Yes, but smallest share |

Key finding: 90.1% of failures are fixable at the interface layer.

Four-Layer Lifecycle: An Exoskeleton for Agents

1. Environment Contract layer (before interaction): enriched tool descriptions, interface constraints, common pitfalls. Removing it drops Airline accuracy by 8.3% and Retail by 17.5%.

2. Procedural Skill layer (task-conditional): BM25 retrieval of reusable procedural skills distilled from training trajectories. Removing it drops Retail by 15.9% and Telecom by 17.4%.

3. Action Realization layer (after generation, before execution): the "final security check" — validates tool names, JSON syntax, required parameters, types, and blocks dangerous patterns. The most critical layer: removing it drops Airline by 61.7% and OS by 59.6%.

4. Trajectory Regulation layer (after environment feedback): detects repeated actions, state oscillation, budget exhaustion, invalid retries, with a graded response (normal → soft recovery → warning → forced correction). Removing it causes a devastating 86.5% drop on ALFWorld.

Results: 18 Models, 116 of 126 Settings Improved

| Benchmark | Environment | Without | With | Gain | |---|---|---|---|---| | AgentBench | ALFWorld | 41.1% | 75.7% | +84% | | | WebShop | 31.4% | 44.0% | +40% | | | OS | 34.7% | 41.2% | +19% | | | DBBench | 48.4% | 64.6% | +34% | | tau-bench | Airline | 49.7% | 62.6% | +26% | | tau2-bench | Telecom | 55.3% | 69.0% | +25% |

Overall: 116 improvements across 126 settings, average relative gain of 88.5%.

The most striking result is cross-model transfer: the harness is evolved only from Qwen3-4B-Instruct trajectories, then applied directly to 17 other models (Qwen 2.5–3.6 series, Llama 3.1-8B/3.3-70B, xLAM 2-3B/8B/32B/70B). All improve — evidence that the harness captures environment-side structure, not model-specific behavior.

Harnessing > Training

Compared to xLAM-2-32B (purpose-trained for tool use):

| Scenario | Qwen2.5-32B + Harness | xLAM-2-32B (tool-trained) | |---|---|---| | tau-bench (in-domain) | 57.9% | 50.4% | | tau2-bench (OOD) | 79.2% | 45.8% |

And adding the harness to xLAM itself yields a further 6.8–28.9 point gain. Training parameterizes environment knowledge; harnessing exposes it structurally. They are complementary, but the harness's OOD transfer far exceeds training.

Evolution Mechanism

The harness is not hand-written — it is evolved:

1. Freeze the model, run training tasks, collect full trajectories 2. A coding agent (Codex) reads trajectories + current harness + design principles 3. Diagnose recurring failures, locate the earliest actionable lifecycle point 4. Propose targeted updates with regression checks against over-triggering 5. Validate safety on unseen tasks

The test set stays hidden throughout; the harness evolves only on training data and is frozen after evolution — ensuring auditability and preventing overfitting.

Limitations

  • Deterministic environments only: stable tool interfaces, clear feedback rules, fixed evaluation criteria. Open-ended tasks have no definable interface contract.
  • Needs execution trajectories to evolve the harness for new environments.
  • Layer dependence varies by task: ALFWorld relies heavily on trajectory regulation, OS on action realization.
  • Evolution quality depends on the coding agent's diagnosis.
  • Gains scale inversely with model size: small models benefit most; large models see smaller relative improvements.

The Bigger Picture: Layered Trust

LIFE-HARNESS signals a shift from "single-layer trust" (the model does everything) to layered trust: the model handles reasoning (black box), the harness handles interface calibration and monitoring (white box, deterministic, auditable), and the environment executes deterministic rules. For enterprise deployment, this is a form of structured runtime safety guardrail — far stronger than a prompt-level "please follow safety policy."

Paper: *Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents* — Tianshi Xu, Huifeng Wen, Meng Li (Peking University), arXiv:2605.22166

Tags

#life-harness#llm-agents#agent-interface#harness-adaptation#deterministic-environments#runtime-intervention#peking-university#model-agnostic

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980543