What’s changed: from static harnesses to self-evolving agents
Previously, the common belief was that an LLM agent could recover from a failure within a single episode or after a retry, but the same mistake would often recur in later tasks because the agent's underlying harness—the tools, context, memory, and workflow—remained unchanged after deployment. This is exactly the problem Living-Harness addresses: it turns each completed trajectory and its evaluator feedback into evidence for updating the harness itself, not just the current episode [1]. This is a shift from static reliability (fixed tools and workflows) to dynamic improvement, where the agent's persistent knowledge accumulates across tasks.
The key innovation is that Living-Harness writes two types of procedural knowledge: episodic memory (recording trigger conditions, failure patterns, and recovery actions) and a state graph (recording state nodes, repair edges, and transition rules). These are retrieved to guide future interactions, while the base tools and context remain frozen. In tests across eight interactive environments, this approach improved average Pass@1 (the success rate on the first attempt) by 10.07 and 9.91 percentage points over the strongest interactive baseline, respectively [1]. That's a substantial gain, showing that updating the harness can directly reduce repeated mistakes.
How hierarchy helps: breaking tasks into levels to reduce error
Hierarchy itself is a powerful tool for long-horizon tasks because it decomposes a complex problem into manageable levels, each with its own decision-making scope. HiSOMA, for instance, uses a three-level structure: a top-level controller learns high-level decision rules, middle-level controllers handle sub-goals, and bottom-level controllers execute primitive actions. This modular design allows the system to handle long-horizon, multi-agent, multi-task problems more efficiently than non-hierarchical approaches, as demonstrated in experiments [2]. The hierarchy helps by reducing the action space at each level, making it easier to learn correct policies and avoid errors that come from trying to decide everything at once.
Similarly, the multi-hop knowledge graph reasoning framework decomposes reasoning into two levels: a high-level relation detector and a low-level entity reasoner. This decomposition effectively controls the action space and improves the rationality of reasoning, leading to superior performance over state-of-the-art baselines on four benchmark datasets [3]. While these papers don't explicitly measure mistake repetition, the underlying principle—that hierarchical decomposition reduces complexity and improves decision quality—directly supports the idea that such structures can help avoid repeated errors in long tasks.
What’s still missing: the catch and open questions
The evidence is promising, but there are important caveats. First, the Living-Harness study is the only one here that directly addresses mistake repetition, and it does so in specific interactive environments (τ²-Bench and MultiWOZ-2.4). It's not clear how well this generalizes to other domains or task types. Second, the hierarchical models (HiSOMA and the knowledge graph framework) show efficiency gains, but they don't explicitly track or prevent repeated mistakes—they just reduce the likelihood of errors by better structuring decisions. So the answer is nuanced: hierarchy helps, but without an explicit memory of past failures, it may not fully prevent repetition.
Another limitation is that the Living-Harness approach requires evaluator signals (feedback) to update the harness. In many real-world tasks, such feedback may be sparse or noisy, which could limit the effectiveness of self-improvement. Also, the study notes that the evolved harness state can be reused across different model backbones, but it doesn't address potential risks like overfitting to past failures or the cost of maintaining a growing state graph. These are open questions that future research will need to tackle.
About These Sources
This answer is built on 3 studies (2 peer-reviewed, 1 preprint) — published from 2022 to 2026, 2 from 2024 or later, 2 in Q1 journals, collectively cited 56 times — selected as the most relevant from 3 studies that passed quality screening, drawn from 41 papers retrieved from a database of over 500 million.
Sources used in this answer
Living-Harness Is an Interactive-Agent Evolver
Living-Harness, a self-evolving agent harness, converts completed trajectories and evaluator signals into procedural memory and a state graph, improving average Pass@1 by 10.07 and 9.91 percentage points over the strongest interactive baseline across eight environments from τ²-Bench and MultiWOZ-2.4.
HiSOMA: A hierarchical multi-agent model integrating self-organizing neural networks with multi-agent deep reinforcement learning
HiSOMA, a hierarchical multi-agent model with a top-level self-organizing neural network and lower-level MADRL controllers, outperforms non-hierarchical MADRL approaches in long-horizon, multi-agent, multi-task problems, demonstrating efficiency gains across experiments.
Step by step: A hierarchical framework for multi-hop knowledge graph reasoning with reinforcement learning
A hierarchical reinforcement learning framework for multi-hop knowledge graph reasoning, decomposing tasks into relation detection and entity reasoning levels, outperforms state-of-the-art baselines on four benchmark datasets, improving action space control and reasoning interpretability.