HGM-Dual: Lineage-Aware Harness for Planning Agents
Harness optimization improves agents built on frozen or closed-source LLMs by refining prompts, tools and workflows. Darwin-Godel Machine and Huxley-Godel Machine rely on recursive self-improvement, mostly for coding tasks. Planning tasks need several interacting constraints met at once, and open-ended self-modification often retries useless edits and wastes the evaluation budget. This paper proposes HGM-Dual with two designs. First, it represents the evolution as lineage-aware context that records ancestor edits, behavior gains and regressions, and ineffective changes. Second, it generates dual-path children based on errors: one general child keeps exploration open, while several specialized children target the components tied to the most violated constraints. HGM-Dual raises case accuracy by up to 14.45% over the strongest baseline and matches the final performance of HGM variants with 2 to 4 times less evaluation budget.
Background and Context
This paper, posted to arXiv, tackles harness optimization for planning agents. A harness is the layer of prompts, tools and workflows built around a frozen or closed-source large language model. Teams cannot change the model weights in that setting. The only lever left is the shell around the model, so harness optimization has become a practical route to a stronger agent.
The best-known methods here are Darwin-Godel Machine and Huxley-Godel Machine (HGM). Both rely on recursive self-improvement: the agent repeatedly edits itself and keeps what works. They were developed mostly on coding tasks. In coding there is a useful coincidence. The skill needed to improve the agent is the same skill the agent is being improved at, so the editing loop and the target task line up.
Planning tasks break that alignment. A planning agent must produce a sequence of tool calls, and several interacting constraints must hold at the same time. The authors argue that open-ended self-modification performs poorly in this setting. It keeps trying edits that have no effect, and each failed attempt burns evaluation budget. Their answer is HGM-Dual, a framework guided by diagnostic information, built to make harness evolution cheaper for planning agents.
Deep Analysis
HGM-Dual has two core designs. The first is lineage-aware context. A common approach stores past agents as a flat archive and lets the editor dig through it. HGM-Dual instead records, for each ancestor, what was changed, which behaviors improved, which behaviors regressed, and which edits did nothing. The editor therefore sees the link between a change and its result. It can tell which kinds of edits help and which are wasted effort, and it avoids retrying the same dead end. The second design is error-conditioned dual-path child generation. For each selected parent, the system produces two kinds of children. The general child keeps exploration open and does not restrict the direction of the edit. This preserves the chance of finding an unexpected gain. The specialized children target the constraints that the parent violated most often. The system first judges which harness components most likely cause those violations, then edits only those components.
How does HGM-Dual decide where to edit a planning harness? It reads the diagnostic signal from the parent's own failed cases. It ranks the most violated constraints and maps them to the harness components most likely responsible. The editor does not guess where to intervene. This keeps the limited evaluation budget on the edits most likely to pay off, while the general child guards against searching too narrowly. The paper's core metric is case accuracy: the share of planning cases in which all constraints are satisfied at once. The comparison set includes the strongest baseline and several HGM variants. The abstract reports two results. HGM-Dual raises case accuracy by up to 14.45% over the strongest baseline. It also matches the final performance of the HGM variants with 2 to 4 times less evaluation budget. The phrase "up to" matters. It means the gain depends on the setting, and 14.45% is the best case, not a guaranteed figure. The abstract also does not name the benchmarks or the underlying models. It does not describe the ablations either. How much each design contributes, lineage context or dual-path generation, cannot be settled from the abstract alone. Readers should check the full paper before drawing that conclusion.
Industry Impact
What does this mean for teams that build agents on closed-source models? The paper speaks to a concrete question. When the model is fixed, can the tuning of prompts, tools and workflows be done in a more structured and cheaper way? The authors say that recursive self-improvement, which works for coding, cannot be moved unchanged to planning, because the two tasks no longer share the same required skill. Engineering teams should therefore look at the nature of the task before choosing an automated optimizer.
The budget result carries real weight. Scoring a candidate harness on planning tasks usually means running a whole batch of cases end to end. Each candidate is expensive. A method that reaches the same final accuracy with 2 to 4 times less evaluation budget can turn the savings into more evolution rounds or a lower bill. For groups that pay per call to a closed model, that difference is direct.
The lineage idea has value beyond this paper. Giving an editor an attributed history of changes is more useful than handing it a pile of old versions. The dual-path idea offers a plain way to balance exploration and exploitation: keep some children open and aim others at the loudest errors. Scenarios with several coupled constraints, such as itinerary building, resource scheduling and customer service flows, are natural places to test these ideas.
Outlook
Several open questions follow from the abstract. The diagnostic signal could become finer. Today it points at the most violated constraints and their related components. A more detailed signal might separate causes inside a single component. The lineage record also grows as evolution continues. Whether it can be compressed without losing the attribution that makes it useful is not addressed in the material available here.
A second question is horizon. The abstract frames the work around planning tasks with interacting constraints, but it does not say how the method behaves on much longer tasks. It does not say how it behaves when constraints are harder to detect automatically. Those claims would need evidence from the full paper and from independent replication.
For practitioners, the sensible next step is to read the paper's benchmark list, model choices and ablation tables. The headline numbers, up to 14.45% higher case accuracy and a 2 to 4 times smaller evaluation budget, are strong signals. They are still reported in an abstract. Teams should confirm that the test setting resembles their own planning workload before adopting the method, and should measure the budget savings on their own cases.
Sources
FAQ
What is HGM-Dual?
HGM-Dual is a method from an arXiv paper that automatically optimizes the prompts, tools and workflows around frozen or closed-source large language models for planning tasks. It turns evolution history into lineage-aware context and generates one general child plus several specialized children based on errors.
Why does HGM-Dual matter for planning agents?
Planning tasks require several interacting constraints to hold at once, and open-ended self-modification keeps retrying useless edits that waste evaluation budget. HGM-Dual raises case accuracy by up to 14.45% over the strongest baseline and matches HGM variants' final performance with 2 to 4 times less evaluation budget.
What should readers check before relying on HGM-Dual's results?
The summary does not name the benchmarks, the underlying models or the ablation details, so the separate contribution of lineage-aware context and dual-path generation remains unclear. Readers should check the full paper for those details, including how each part contributes, before drawing conclusions about HGM-Dual.