Source-linked AI summary
HGM-Dual: Lineage-Aware Guided Agent Harness Optimization for Long-Horizon Planning
Sudipta Paul, Pengrui Quan, Aounon Kumar, Nicholas Tzou, Vivek Kulkarni, Srinivas Chappidi
TL;DR
Planning tasks require sequences of actions that satisfy interacting constraints, challenging recursive self-improvement approaches developed mainly for coding. HGM-Dual combines structured lineage-aware edit outcomes with error-conditioned general and specialized child generation. On travel and shopping planning tasks, it outperforms the strongest baseline and matches HGM variants’ final performance with fewer evaluations.
Problem
Planning requires action sequences that satisfy interacting constraints, while coding-oriented recursive self-improvement benefits from alignment between the competence being improved and that needed to improve it.
Source: §1 · p. 1Method
HGM-Dual records lineage-level edit outcomes and generates both a general exploratory child and specialized children targeting harness components implicated by parent errors.
Source: §1 · p. 2Results
HGM-Dual achieves up to 14.45% relative improvement in case accuracy over the strongest baseline while matching HGM variants’ final performance with 2×–4× fewer evaluation budget.
Source: §1 · p. 2Takeaways & Limitations
Experiments on travel and shopping planning tasks show that structured evolutionary context and diagnostic guidance can improve harness evolution while reducing evaluation budget.
Source: §1 · p. 2In HGM, the archive does not explicitly represent structured edit–outcome evidence, and open-ended edits are not explicitly constrained to address the selected parent’s failures.
Source: §3.2 · p. 4
Abstract
from arXiv · showhide
Harness optimization, which improves an agent's prompts, tools, and workflows, is an increasingly effective way to enhance agents built on frozen or closed-source LLMs. Leading approaches such as the Darwin--Godel Machine and the Huxley--Godel Machine (HGM) rely on recursive self-improvement and were developed mainly for coding, where the competence needed to improve the agent coincides with the competence being improved. Planning tasks, in which an agent must produce a sequence of tool calls that jointly satisfies interacting constraints, break this alignment. There, open-ended self-modification repeatedly explores ineffective changes and consumes substantial evaluation budgets. We introduce HGM-Dual, a diagnostic-guided framework for budget-efficient harness evolution for planning agents. First, HGM-Dual represents agent evolution as lineage-aware context. This context records what each ancestor changed, which behaviors improved or regressed, and which modifications were ineffective, giving the editor explicit edit-outcome attribution instead of an unstructured archive of past agents. Second, HGM-Dual introduces error-conditioned dual-path child generation. For each selected parent, it generates a general child that preserves open-ended exploration and specialized children that target the harness components most likely responsible for the parent's highest-ranked constraint violations. HGM-Dual outperforms the strongest baseline by up to 14.45% relative improvement in case accuracy while matching HGM variants' final performance with 2x-4x fewer evaluation budget.
1 INTRODUCTION
Planning harness optimization is difficult because constraint-dependent action sequences break the alignment between the competence being improved and the competence needed for self-improvement. HGM-Dual combines lineage-aware edit outcomes with error-conditioned exploration and reports stronger results at lower evaluation cost.
1 INTRODUCTION
- Planning breaks the coding-oriented alignment of recursive self-improvement, while HGM’s shopping search repeats prior edit patterns in 41.8% of child nodes.Repeated attempts consume another full planning evaluation, motivating explicit edit–outcome context.Source: §1 · pp. 1–2, Figure 1 · p. 2
- HGM-Dual gives the editor structured lineage context recording changes, behavioral improvements or regressions, and ineffective modifications.This provides explicit edit–outcome attribution rather than an unstructured archive of agent snapshots.Source: §1 · p. 2
- For each parent, HGM-Dual generates a general child for open-ended exploration and specialized children targeting harness components implicated by observed errors.The two paths balance exploration with diagnostic-guided local refinement.Source: §1 · p. 2
- HGM-Dual achieves up to 14.45% relative improvement in case accuracy over the strongest baseline while matching HGM variants’ final performance with 2×–4× fewer evaluation budget.Experiments cover travel and shopping planning tasks.Source: §1 · p. 2
2 RELATED WORKS
Prior work improves frozen LLM agents by modifying their harnesses, including through iterative and evolutionary self-modification. The section identifies gaps in how prior edits inform generation and how edits target observed failures.
2 RELATED WORKS
- Harness optimization methods modify prompts, tools, skills, or workflows, and approaches such as ADAS, DGM, HGM, GEPA, and HyperAgents search agent implementations through iterative or evolutionary self-modification.Meta-Harness uses prior candidates’ code, scores, and traces; SkillOpt uses bounded edits and validation.Source: §2 · pp. 2–3
- Prior methods either provide local or excessive edit experience, which can obscure useful edit–outcome relationships and lead to repeated ineffective modifications across the search tree.DGM and HGM primarily use the selected parent’s code and logs, whereas archive-based methods may supply numerous candidates and raw traces.Source: §2 · p. 3
- Prior methods guide candidate selection or validation, but edit generation remains largely open-ended and may miss the harness component responsible for observed failures.HGM-Dual addresses this with structured evolution context and both general and error-conditioned specialized children.Source: §2 · p. 3
3 HGM-DUAL
HGM-Dual evolves frozen-model harnesses by giving editors structured lineage history and pairing open-ended edits with error-targeted edits. Its expansion process retains the strongest child while discarding the others.
3.1 PROBLEM SETUP
- Harnesses comprise a prompt, available tools, and an agentic workflow, while the underlying language model remains fixed during optimization.The objective is to maximize expected task-specific score over the task distribution.Source: §3.1 · p. 3
3.1 PROBLEM SETUP; 3.2 OVERVIEW
- At each iteration, HGM selects a harness from its archive, edits it, and evaluates the child; HGM-Dual adds lineage context and error-conditioned editing to this process.The setup treats the model parameters as fixed and optimizes the harness over expected task performance.Source: §3.2 · pp. 3–4, §3.1 · p. 3
3.3 AGENT EVOLUTION AS INTERPRETABLE CONTEXT
- HGM-Dual summarizes ancestral edits and outcomes into a bounded lineage context, helping editors avoid failed changes and notice interactions not evident from the current score alone.The records connect modifications with functionality, performance impact, and behavioral consequences.Source: §3.3 · pp. 4–5
3.4 ERROR-CONDITIONED DUAL-PATH CHILD GENERATION
- The general path explores broadly, while specialized children target the parent’s highest-frequency error categories with focused corrections to prompts, tools, or workflows.Error-conditioned editing is enabled only after a pre-search test meets κ ≥0.4 and ρB ≥0.7; otherwise, expansion uses only the general path.Source: §3.4 · pp. 5–6
- HGM-Dual adds the best child, its evaluation context, and edit attribution to the archive, discarding the other children to limit archive growth.All generated children are evaluated on the same optimization set with the same reward function.Source: §3.4 · p. 6
4 EXPERIMENTS
The experiments evaluate harness optimization on constrained travel and shopping planning tasks, using held-out instances and a fixed evaluation budget. HGM-Dual improves on the strongest baseline, with results also examining budget efficiency, failure categories, component contributions, and a second task-agent model.
4 EXPERIMENTS
- The benchmark tests globally constrained travel itineraries and shopping carts; evaluation measures task performance and counts plans generated during optimization as the compute budget.Travel uses nine APIs and 21 feasibility checks; shopping searches an average of 171 records through 15 APIs. Optimization uses 60 samples and evaluation uses 120, including 60 unseen samples; model weights are not updated.Source: §4 · pp. 6–7
4.1 EVALUATION
- Figure 4 tracks performance across 1k–4k evaluations and failure-rate changes on 60 held-out instances; HGM-Dual improves each travel checkpoint and reaches 90.4 shopping composite at 1k.HGM-CF stalls on travel after 1k, while HGM-Dual reaches its final shopping score using roughly a quarter of HGM-CF’s budget.Source: Figure 4 · p. 7, §4.1 · p. 8
- HGM-Dual reduces travel tasks with any error to 34.6%, versus 47.5% for HGM-CF, and cuts shopping unmet-feature failures from 24.2% to 10.8%.On travel, it also reduces route-consistency failures from 16.7% to 1.7% and sandbox-compliance failures from 10.0% to 2.5%.Source: Figure 4 · p. 7, §4.1 · p. 8
- At 1,000 evaluations, HGM-Dual leads shopping with 90.35 composite and 70.83 case accuracy, exceeding the strongest competitor by 5.89 and 14.16 points.HGM-CF is strongest on travel at this budget, with 84.90 composite and 47.50 case accuracy.Source: Table 2 · p. 8, §4.1 · p. 9
- At 4,000 evaluations, HGM-Dual beats strongest baseline HGM-CF by 3.66/2.56 composite and 8.13/5.84 case-accuracy points on travel/shopping.The gains are larger in case accuracy than composite score in both domains.Source: §4.1 · p. 8
- Adding Agent Evolution Context lowers lineage edit redundancy from 0.22 to 0.17, reduces edits worsening instances from 37.6% to 31.9%, and raises child improvements from 31.3% to 41.0%.The comparison is between HGM-CF and HGM-CF with Agent Evolution Context, without error-conditioned child generation.Source: Table 3 · p. 9, §4.1 · p. 9
- At 1,000 evaluations, the error-conditioning ablation scores 87.68 versus 90.35 for HGM-Dual; adding evolution context and then error-conditioned generation further improves performance.The ablation keeps the number and sequence of children and best-child selection unchanged, omitting the error category from the generation prompt.Source: Table 4 · p. 9, §4.1 · p. 9
- With Mistral-Medium-3.5-128B as task agent, HGM-Dual reaches composite score 42, 10 points above seed and 8 above HGM-CF, while retaining seed case accuracy of 1.6%.HGM-CF raises composite score from 32 to 34 but reduces case accuracy from 1.6% to 0.8%.Source: Table 5 · p. 9, §4.1 · p. 9
5 CONCLUSION
HGM-Dual combines structured agent-evolution context with error-conditioned dual-path generation for more effective, resource-efficient harness evolution. The conclusion also identifies broader lineage sharing, adaptive management, and wider evaluation as future work.
5 CONCLUSION
- HGM-Dual’s complementary mechanisms outperform strong self-improving baselines while substantially reducing the evaluation budget.The conclusion attributes this approach to explicitly modeling evolutionary history and balancing general with error-targeted search.Source: §5 · p. 10
- Ablations show that evolution context provides consistent performance gains, while dual-path generation accelerates improvement under constrained budgets.Source: §5 · p. 10
- Future work could extend edit-outcome sharing across the search tree, jointly learn context and dual-path management, and evaluate broader domains and online environments.The paper states these evaluations would help characterize transferability and long-term robustness.Source: §5 · p. 10
AI USE STATEMENT
The authors report that they used generative AI for drafting and editing prose, while the methodology was designed and implemented by the authors. They state that they reviewed AI-assisted work and take responsibility for the final content.
AI USE STATEMENT
- The authors say all research ideas, methods, experiments, and findings are their work.Source: AI USE STATEMENT · p. 10
- Generative AI tools helped draft text from author-provided materials and edit grammar, clarity, and readability.Source: AI USE STATEMENT · p. 10
- The authors state that they checked technical statements, numbers, and citations, revised AI-drafted text, and take responsibility for the final content.Source: AI USE STATEMENT · p. 10
A.1 EDIT REDUNDANCY
The edit-redundancy analysis tags proposals by functional error categories and compares exact category combinations across the HGM and HGM-Dual search-tree archives.
A.1 EDIT REDUNDANCY
- The analysis examines non-root edit proposals in the shopping search-tree archives, with 67 HGM proposals and 51 HGM-Dual proposals.Source: §A.1 · p. 12
- Proposals are tagged using seven error categories, including coupon correctness, profile compliance, shipping constraints, budget compliance, cart completeness, exact feature matching, and price optimality.Source: §A.1 · p. 12
- Because an edit may address multiple errors, its edit pattern is the exact combination of its assigned categories; patterns are ranked separately for HGM and HGM-Dual.Two proposals count as the same pattern only when their complete category sets are identical.Source: §A.1 · p. 12
- The figure reports the share of repeated edit patterns in each method’s archive.Source: §A.1 · p. 12
A.2 EDIT-TARGETING MISMATCH
The targeting-mismatch analysis maps each edit’s stated objective and proposed changes to functional error categories and checks whether targets correspond to errors reported for the parent.
A.2 EDIT-TARGETING MISMATCH
- Each non-initial candidate is treated as an edit and compared with the candidate from which it was derived.Source: §A.2 · p. 12
- An edit has no corresponding parent error category if it targets at least one functional component for which the parent reported no error.Source: §A.2 · p. 12
- The dominant error category is the category associated with the largest number of failing evaluation cases.Source: §A.2 · p. 12
A.3 ESTIMATED COST
The section estimates task-agent API costs for 1,000-sample Travel and Shopping evaluations and explains how evaluation costs shape the experimental budget.
A.3 ESTIMATED COST
- After harness optimization, 1k case evaluations can cost hundreds of dollars in task-agent API usage.The estimate motivates prioritizing the available budget toward independent runs of HGM variants.Source: §A.3 · p. 12
- ADAS and GEPA are evaluated at the 1k budget, while HGM, its variants, and HGM-Dual receive the more expensive 4k experiments.The authors prioritize independent runs of HGM variants because they are the study’s primary focus.Source: §A.3 · p. 12
- 1,000-sample Travel and Shopping evaluation costs are estimated for optimized averaged candidate agents evaluated by HGM and HGM-Dual.The task agent is GPT-5.4-mini at high reasoning effort; the estimates give lower and upper bounds based on prompt caching.Source: Table 6 · p. 13
A.4 FULL HGM-DUAL
HGM-Dual retains HGM’s search loop but changes how selected parents are expanded. Thompson sampling allocates expansions and evaluations, and the final harness is selected using a posterior percentile.
A.4 FULL HGM-DUAL
- At each iteration, HGM-Dual either expands a selected parent or evaluates an existing harness on another task, then adds only the strongest generated child to the archive.Expansion uses lineage context to generate one general child and two error-conditioned children.Source: §A.4 · p. 13
- Thompson sampling selects both the parent to expand and the harness to evaluate, and the search returns the harness with the highest posterior success-rate percentile.The final selection uses Iρ, the ρ-percentile of a harness’s Beta posterior.Source: §A.4 · p. 13
- Clade-level passed and failed task counts guide parent selection, while each harness’s own counts guide evaluation and final selection.Clade counts aggregate over the subtree rooted at a harness.Source: §A.4 · p. 13
- The scheduler sharpens sampled Beta distributions over time, shifting search from exploration toward exploitation.The widening parameter α controls how quickly the archive grows relative to the evaluation budget.Source: §A.4 · p. 13
A.5 AGENT EVOLUTION CONTEXT
The appendix illustrates lineage-context records and observed outcomes for travel and shopping harness changes, then specifies editor instructions for diagnosing and instrumenting modifications.
A.5 AGENT EVOLUTION CONTEXT
- For travel, the harness added a selector that parses and ranks intercity options, while the workflow prompt requires an explicit selected transport line on transfer days.The selector is exposed as a callable tool.Source: §A.5 · p. 13
- The intercity selector was used 122 times with no observed errors, yet appeared in all 26 passing and 34 failing cases, so usage alone did not distinguish outcomes.The instrumentation logged selector-mode and verifier events alongside selector use.Source: §A.5 · p. 13
- Travel instrumentation verdicts did not reliably track case outcomes, and a transport parse failure occurred in a case that nevertheless passed.Case 110 failed with verifier_fired/pass, while several passing cases had verifier_fired/fail.Source: §A.5 · p. 14
- Travel failures included transfer-time and transport correctness issues as well as accommodation, restaurant, and budget errors.The selector remained compatible with successful runs, but transport-focused cases also succeeded with the tool active.Source: §A.5 · p. 14
- Shopping changes included cart reset, prefetched user and cart context, and level-specific planning and verification checklists.The appendix also describes instrumentation for those prefetch and checklist steps.Source: §A.5 · pp. 14–15
- Shopping performance reached 37/60 observed cases; prefetch and checklist checkpoints fired in all 12 new cases, which included 8 passes and 4 failures.Across the observed cases, those paths appeared in all 60 cases.Source: §A.5 · p. 15
- The shopping checklist and prefetch instrumentation also fired in every failing case, where failures mostly involved cart mismatches or coupon handling.Coupon handling included 11 calls with 4 errors.Source: §A.5 · p. 15
A.6 AGENT EDITOR INSTRUCTION
- The editor is instructed to inspect the task agent and grading materials, target score-relevant failures, and modify only specified workspace files under import and tool-routing rules.When adding a verifier, helper, or decision branch, it must log a labeled pass, fail, or skip verdict at the decision point, except for trivial edits.Source: §A.6 · pp. 15–16