Living-Harness Is an Interactive-Agent Evolver

Abstract: LLM agents may recover from a failure within an episode or after a retry, yet the same execution failure can recur in later tasks because post-episode feedback rarely revises the persistent harness that guides future interactions. Static harnesses improve reliability through fixed tools, context, memory, and workflow structures, but remain unchanged after deployment. We propose $\textbf{Living-Harness}$, a self-evolving agent harness that converts each completed trajectory and its evaluator signals into posterior evidence for bounded harness updates. Guided by a domain-level $\textbf{Evolution-SOP}$ ($\textbf{S}$tandard $\textbf{O}$perating $\textbf{P}$rocedure), Living-Harness extracts an episode abstraction and structured update evidence, and writes two complementary forms of procedural knowledge: episodic memory that records trigger conditions, failure patterns, and recovery actions, and a state graph that records state nodes, repair edges, and transition rules. The updated harness state is retrieved to guide future interactions, while tools and base context remain frozen, allowing procedural repairs to accumulate across evolution cycles. On eight interactive environments derived from $τ<sup>2$-Bench and MultiWOZ-2.4, Living-Harness improves average Pass@1 over the strongest interactive baseline by 10.07 and 9.91 percentage points, respectively, and supports retrieval-only reuse of the evolved harness state across model backbones.

• The paper introduces a self-evolving harness that converts evaluator-grounded trajectories into gated updates to episodic memory and state graphs while freezing the base model, tools, and core context.
• Living-Harness achieves average Pass@1 scores of 83.09 on τ²-Bench and 65.50 on MultiWOZ-2.4, outperforming the strongest interactive baselines by 10.07 and 9.91 percentage points.
• The method supports cross-model procedural transfer and substantial recovery gains, but remains limited by simulator-based testing, manually defined evolution rules, added computation, and the absence of rollback safeguards.

Living-Harness addresses a specific deficiency in LLM agent systems: the gap between task-local correction and cross-episode procedural repair. An agent may recognize a failure within an episode or recover on retry, yet the same failure recurs in later tasks because post-episode feedback rarely revises the persistent harness—the external layer of prompts, tools, memory, and workflow structure that organizes the base model's behavior. The paper proposes Living-Harness (2607.26598), a self-evolving harness that converts evaluated trajectories into bounded updates of two persistent structures—episodic memory and a state graph—while keeping tools, base context, and the underlying model frozen. On eight interactive environments derived from $\tau^2$-Bench and MultiWOZ-2.4, it improves average Pass@1 over the strongest interactive baseline by 10.07 and 9.91 percentage points respectively.

Motivation and problem formulation

The authors frame the target capability as persistent procedural repair: converting evaluator-grounded failure evidence into a reusable relation between when a failure occurs, what action or transition is missing, and how a later rollout should recover. They argue that existing work covers only parts of this. Harness-design and workflow-optimization methods (e.g., AFlow, Meta-Harness) build strong procedural structure but freeze it before deployment; reflection and experience-memory methods (Reflexion, ExpeL, ReasoningBank) retain critiques or trajectories, but a textual lesson such as "transfer the user to a human" does not specify the trigger condition, required tool call, and workflow transition needed to prevent recurrence. Living-Harness requires four properties: grounding in completed trajectories plus evaluator signals (not unverified self-critique), persistence across episodes with task-family scoping, modification only of stateful harness components while tools and base context remain frozen, and capture of both experiential knowledge (why a failure occurred) and workflow knowledge (which transition should change).

The framework is formalized through a program-state POMDP lens. Beyond the standard environment-state belief $b_t(s)$, the agent maintains an episode-level program state $z^{(n)}$ induced by the harness state $S^{(n)}$, yielding an augmented belief $B_t^{(n)}(s,z)$ over the joint $(s_t, z^{(n)})$. Proposition 1 establishes that for nested information sets $\mathcal{I}_s \subseteq \mathcal{I}_{s,z}$, the Bayes error bound satisfies $$\mathcal{B}(\mathcal{I}_{s,z}) = \mathcal{B}(\mathcal{I}_s) - \Gamma_z \leq \mathcal{B}(\mathcal{I}_s)$$, with strict improvement whenever $z$ carries nonzero conditional information about execution-relevant latents. The supplementary material extends this to approximate estimators (the gain holds when $\epsilon_z - \epsilon_s < \Gamma_z$, i.e., information gain exceeds approximation error) and to general decision risk. This is an information-theoretic justification, not a performance guarantee: strict improvement depends on $b_t(s)$0 actually being informative, which is an empirical property of the retrieval and extraction pipeline.

Method

The harness at episode $b_t(s)$1 is $b_t(s)$2, where $b_t(s)$3 contains frozen tools, base context, and domain rules; $b_t(s)$4 is a fixed domain-level Evolution-SOP; and $b_t(s)$5 is the evolving state comprising episodic memory and a state graph. Initialization provides no memory and only a coarse graph scaffold; all fine-grained repairs are induced from evaluated experience.

The rollout–evaluate–update loop proceeds in three stages. First, posterior abstraction: given task $b_t(s)$6, trajectory $b_t(s)$7, evaluator signals $b_t(s)$8, and fixed resources, the Evolution-SOP produces an episode abstraction identifying the objective, verified facts, outcome, and critical failure or recovery point. Second, evidence extraction: the abstraction is decomposed into memory evidence $b_t(s)$9 (trigger conditions, failure patterns, recovery actions) and graph evidence $z^{(n)}$0 (states, actions, transitions). Third, gated commitment: candidates are checked against five gates—schema, scope (task-family isolation), evidence (grounding in evaluator feedback or repeated failures), constraint (no override of frozen policies or tool preconditions), and merge (semantic deduplication with confidence accumulation). Unsupported candidates leave the state unchanged. Crucially, the updated state becomes available only to subsequent episodes; the paper enforces a score-before-update protocol so that evidence from an episode never influences that episode's reported score. Task-local reflexion (up to three trials per instance) is kept in a local buffer and never written into global state, cleanly separating within-task correction from persistent evolution.

At rollout time, a task-conditioned query retrieves top-$z^{(n)}$1 entries from both components under a same-family-first strategy, rendering them as actor-facing procedural context while fixed rules retain precedence.

Experimental results

All GPT-5.2-based methods use GPT-5.2 (medium reasoning effort) as backbone; the simulated user on $z^{(n)}$2-Bench uses GPT-5.1. Baselines comprise flagship models evaluated without online evolution (Gemini 3 Pro, GLM-5, Qwen3-max, GPT-5.2, Kimi-k2) and interactive/self-improving methods (ReAct, Reflexion, AWM, ReasoningBank, EvoTest).

Method	$z^{(n)}$3-Bench avg	MultiWOZ-2.4 avg
Gemini 3 Pro	82.92	55.80
Reflexion (best interactive, $z^{(n)}$4)	73.02	53.10
ReasoningBank (best interactive, MW)	52.52	55.59
Living-Harness	83.09	65.50

Two claims deserve emphasis. First, Living-Harness slightly exceeds the strongest flagship-model average on $z^{(n)}$5-Bench (83.09 vs. Gemini 3 Pro's 82.92) despite using GPT-5.2 medium as its base—a notable result implying that harness-state evolution can substitute for raw backbone capability in these settings. Second, gains concentrate where procedural transfer matters most: on MultiWOZ-2.4, Living-Harness trails Reflexion slightly in the one-domain group but leads decisively in two-domain (70.52 vs. 51.35) and three-domain (25.87 vs. 21.68) settings.

Per-cycle analysis shows recovery accumulation: Retail rises from 57.02 to 85.96 over four cycles, Telecom from 57.39 to 78.07, and MultiWOZ domains improve by up to +28.71 (Taxi). The largest gains appear after the first cycle, with diminishing refinements and occasional mild regressions (e.g., Restaurant drops 1.82 at Cycle 2)—consistent with bounded evolution rather than unbounded context accumulation.

Ablations attribute the largest share to the Evolution-SOP itself: removing it drops the $z^{(n)}$6-Bench average from 83.09 to 73.38, versus 77.34 without memory and 79.50 without the state graph. This supports the claim that structured posterior interpretation and gated commitment—not merely attaching memory or a workflow graph—drive the improvement.

Cross-model transfer is the strongest portability result: a harness state evolved with GPT-5.2 is frozen and provided retrieval-only to Gemini 3 Pro, GLM-5, Qwen3-max, and Kimi-k2, improving every reported domain score. The most striking cases are Taxi, where GLM-5, Qwen3-max, and Kimi-k2 improve from 0.00 to 43.08, 45.13, and 45.13 respectively. Because even Gemini 3 Pro benefits, the evolved state appears to encode backbone-agnostic procedural knowledge rather than compensating solely for weak models. A case study illustrates the mechanism concretely: Reflexion repeatedly identifies that a user should be transferred to a human operator but never executes transfer_to_human_agents(); Living-Harness stores this as both a memory entry and a state-graph repair edge linking terminal-suspension detection to the required tool call, after which the task completes in one attempt.

Limitations and open questions

The paper is explicit about several constraints. Evaluation is confined to simulator-based benchmarks with finite tool sets and evaluator-defined success criteria; robustness to shuffled streams, held-out task families, policy perturbations, and open-ended deployments remains untested. The commit gates reduce malformed or unsupported updates but do not guarantee monotonic improvement—there is no rollback, stale-entry removal, or regression testing, so incorrect evaluator feedback or overly narrow repairs can persist and contaminate future retrieval. Each Evolution-SOP's domain monitoring rules are manually specified; zero-shot transfer of an Evolution-SOP to unseen domains is not established. Finally, post-episode computation adds deployment cost, and persistent interaction histories raise privacy concerns requiring data minimization and retention controls.

Conclusion

Living-Harness reframes interactive-agent adaptation as maintenance of a persistent, retrievable procedure around the model rather than improvement of individual responses. Its empirical contributions—consistent double-digit gains over interactive baselines, cycle-wise recovery accumulation, and retrieval-only reuse across backbones—are supported by ablations showing that the gated Evolution-SOP mechanism, not the memory or graph containers alone, accounts for most of the benefit. The main open questions concern generalization beyond controlled simulators and mechanisms for guaranteeing non-regression of the evolving state.