Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces

Jian Mu; Ruibo Shen; Yao Shu

cs.LGOct 27, 2026

Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces

Authors: Jian Mu, Ruibo Shen, Yao Shu

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology

Abstract

Language model agents are post-trained and deployed through a harness, the prompts, tools, and control logic that present a task to the model and execute its outputs. Harnesses are often revised and judged by how an already trained model performs under them, which assumes that what the model learned under one harness carries over to the next. We ask what a policy learns from its harness and what survives when the harness changes. Under KL-regularized control, we show that the log odds of an optimal policy recover how strongly its harness favors each action over the others, so a trained policy carries the preferences of its harness. Deployed return then equals the ceiling of the harness, the best return it offers, minus the mismatch of the policy, the return lost to its own suboptimal decisions. Training changes only the mismatch, while changing the harness moves both terms. Comparing harnesses with a fixed model therefore mixes what each harness offers with how well that model uses it. In MiniWoB++, post-training makes a model's preferences specific to its training harness. Reversing the order of the observation fields leaves every optimal decision unchanged yet lowers a trained model's success from 85.0% to 19.2%. Conversely, every new harness that a fixed-model comparison rejects but that raises the ceiling outperforms the original agent after further training under it, and ceiling and mismatch measured beforehand predict these gains. For coding-agent harnesses on Terminal-Bench Lite, ceiling minus mismatch estimated with a stronger model as a stand-in for the optimum reproduces the accuracy ranking, while either term alone does not. Code is available at https://github.com/2020-qqtcg/HRP.

Figures & tables

Figure 1: Adaptation reverses a frozen comparison. The left group evaluates the model trained under Harness 0, the source checkpoint, under both harnesses with its weights frozen. The right group evaluates the models after equal further training under each harness. The dotted line marks the source checkpoint’s success under Harness 0, and colors mark the evaluation harness.

Figure 2: A harness revision and a weight update act on different terms. Summit heights are the ceilings, the robot’s altitude is its return, and the gap between them is the mismatch. Moving the trained model to the new harness with frozen weights (① to ②) raises the ceiling but lowers the return. Updating the weights under the new harness (② to ③) lowers only the mismatch. The adapted model beats the source checkpoint under the original harness once its mismatch $M^{\prime}$ falls below $T$ , the new ceiling minus the source checkpoint’s return under the original harness.

Figure 3: Post-training moves action preferences toward the optimum of the training harness and makes them specific to it. Rows denote training harnesses. (a) Task success and (b) geometry error relative to the training harness’s optimum during post-training. Colored curves show the harness of each row, gray curves the other harnesses, and dotted lines the error at initialization. (c) Geometry error of each final checkpoint under each deployment harness. Outlined entries mark the lowest error in each column, attained by the checkpoint trained under that harness.

Presentation	Choice change	$M$	$J$	Task success
Original	0.0%	0.015	$0.423$	85.0%
Reversed fields	52.0%	0.484	$-0.046$	19.2%
Menu as list	9.2%	0.023	$0.415$	86.0%
Shifted codes	27.7%	0.148	$0.289$	57.1%
Reversed codes	30.0%	0.130	$0.308$	68.2%
Permuted codes	25.3%	0.110	$0.328$	67.6%

Table 1: Equivalent optimal decisions, different choices. Choice change is the share of states whose most probable action changes from the original.

M

and

J

are the mismatch and regularized return of Section 5.2 .

Figure 4: Harmful revisions become beneficial after adaptation when they raise the ceiling. (a) Return

J

, ceiling

C

, and mismatch

M

of the H0-trained checkpoint, its H0 continuation, and its H1 adaptation under each harness. The dotted line marks the source pair’s return. Gains of the adapted pair over the source pair in the nine swaps that a frozen comparison rejects and that have

T>0

, plotted against (b) the frozen return change and (c) the predicted gain

T-r_{2}

. In (b) and (c), dark points raise the ceiling, light points lower it, and the red point is H0

\to

H1.

Figure 5: Sampled mismatch tracks the small return differences that decide adaptation. Curves show exact values and the two predictions at each saved H1-adapted checkpoint. (a) Recovery, the checkpoint’s return under H1 minus the source H0 pair’s return. (b) Harness ranking, the checkpoint’s return under H1 minus its return under H0.

Figure 6: Mismatch accumulates over decisions. Each harness’s mismatch splits into the number of decisions the evaluated model makes per trajectory and the mismatch per decision. Dashed curves have constant

\widetilde{M}

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$\mathcal{I}=(c_{\mathcal{I}},d_{\mathcal{I}},e_{\mathcal{I}})$	Context encoder, semantic decoder, and fixed execution protocol.
$\pi_{\theta},\ \Pi_{\theta,\mathcal{I}}$	Policy over token sequences and its executed semantic pushforward.
$P_{\mathcal{I}},r_{\mathcal{I}},\pi^{\mathrm{ref}}_{\mathcal{I}}$	Induced dynamics, task reward including execution costs, and semantic reference prior.
$\rho_{0},\gamma,\tau$	Initial state distribution, discount, and KL regularization coefficient.
$F^{}_{\mathcal{I}},\ \bar{F}^{}_{\mathcal{I}}$	Optimal action score and its centered representative.
$\widehat{F}_{\theta,\mathcal{I}}$	Readout, the centered action scores $\tau P_{\!\perp}\log\Pi_{\theta,\mathcal{I}}$ .

Appendix

Table 3: Notation. Energies and vector operations are statewise unless an expectation is shown.

Panel of Figure 7 — Figure 7: Adaptation can realize an initially harmful interface improvement. Analytical curves for the two-action bandit with $\tau=0.5$ . The frozen policy exceeds the original pair only for $u>e^{2}$ (dashed), while the ceiling gain is positive for $u>1$ . Reaching the new optimum realizes the ceiling gain $C_{2}-C_{1}$ (solid). Corollary 5.2 gives the mismatch threshold for a benefit before full adaptation.

ID	Interface	Actions, execution, and controller state
H0	Direct	One native click or checkbox submission per decision.
H1	Batched clicks	Adds every ordered pair of clicks as a single menu option, executed without an intervening actor call.
H2	Automatic gate	Checks a commit action before execution. Rejection leaves the page unchanged, consumes the decision, and disables subsequent checks. Memory stores whether the gate is enabled.
H3	Verify tool	Adds an optional program check of a proposed native action. Memory stores the checked action and its advisory verdict.
H4	Reviewer tool	Uses H3’s action menu and memory, with advice supplied by a frozen language-model reviewer.
H5	Revised instructions	Uses H0’s actions, observations, dynamics, and budget, with additional system instructions for matching labels, tracking tile IDs, correcting selections, and deciding when to submit.

Appendix

Table 4: MiniWoB++ harnesses. Every menu includes the complete parameters of each action. We measure the common budget in actor decisions.

Split	Per family	Total	Use in post-training
Training	480	1,440	Parameter updates
Calibration	96	288	Full learning curves
Training probe	24	72	Fixed training subset for learning curves
Test	192	576	Transfer evaluation

Appendix

Table 5: MiniWoB++ task allocation. Splits have disjoint public contents, except that the training probe comes from training. Metrics average tasks within each family and then weight the families equally.

Parameter	Setting
Base / reference / H4 reviewer	Qwen3.5-2B with frozen base, reference, and reviewer weights
Trainable adapter	LoRA rank 16, scale $\alpha=32$ , dropout 0, no bias updates
LoRA projection targets	Attention: q/k/v/o_proj Linear attention: in_proj_qkv , in_proj_z , in_proj_b , in_proj_a , out_proj MLP: gate/up/down_proj
Optimizer	AdamW with a constant learning rate $10^{-5}$ , $\beta=(0.9,0.999)$ , $\epsilon=10^{-8}$ , weight decay 0, gradient norm clipping at 1
Training budget	1,024 updates per harness at four tasks per update (4,096 task draws)
Checkpoint schedule	Initialization and every 64 updates (17 checkpoints at the full budget)

Appendix

Table 6: Post-training settings. All harnesses share the same update and task-draw budgets, with computation depending on the number of reachable states and menu entries.

Variant	Readout RMS	Mismatch $M$	Return $J$	Success (%)
R0	0.000	0.015	$0.423$	85.0
	[0.000, 0.000]	[0.010, 0.021]	[ $0.409$ , $0.436$ ]	[84.0, 86.1]
R1	0.200	0.484	$-0.046$	19.2
	[0.198, 0.203]	[0.460, 0.507]	[ $-0.072$ , $-0.020$ ]	[17.0, 21.4]
R2	0.054	0.023	$0.415$	86.0
	[0.053, 0.054]	[0.018, 0.028]	[ $0.401$ , $0.428$ ]	[85.1, 86.9]

Appendix

Table 7: H0 readout results, with 95% paired-task bootstrap intervals below each mean. RMS intervals apply the square root to the mean-squared drift interval endpoints. The target drift is numerically zero in every presentation.

Checkpoint	Deployment	$J$	$M$	Success (%)
Source	H0	0.42289	0.01485	85.05
Source	H1	0.31285	0.13737	72.90
H0 continuation	H0	0.42294	0.01481	83.37
H0 continuation	H1	0.30404	0.14619	75.66
H1 adaptation	H0	0.38642	0.05132	78.31
H1 adaptation	H1	0.43133	0.01889	88.17

Appendix

Table 8: Deployment outcomes of the source checkpoint and both final checkpoints. Both training branches receive 1,024 additional updates.

J

is regularized return,

M

is occupancy-weighted decision loss, and success is the undiscounted probability of completing the task.

Quantity	Mean [95% CI]
Ceiling gain $C_{2}-C_{1}$	$0.01248\ [0.00698,\,0.01754]$
Frozen mismatch increase	$0.12252\ [0.11487,\,0.12961]$
Frozen return change	$-0.11004\ [-0.11558,\,-0.10438]$
Recovery threshold $T$	$0.02733\ [0.01984,\,0.03520]$
Final recovery difference	$0.00844\ [0.00135,\,0.01447]$
Final interface ranking	$0.04491\ [0.03653,\,0.05254]$

Appendix

Table 9: Paired comparisons with 95% task bootstrap intervals. Final comparisons use the H1-adapted checkpoint at 1,024 additional updates. Recovery compares it under H1 with the source H0 pair. Interface ranking compares H1 and H0 at the same adapted checkpoint. Gain over H0 continuation compares the two branches at the same update budget, each under its training harness.

Swap	$\Delta C$	Frozen	$T$	$T-r_{2}$	Gain	Ranking	Success
H1 $\to$ H5	$0.068$	$-0.054$	$0.092$	$0.081$	$0.085$	$0.167$	$+9.0$
H0 $\to$ H5	$0.081$	$-0.004$	$0.095$	$0.085$	$0.085$	$0.111$	$+8.7$
H3 $\to$ H1	$0.078$	$-0.040$	$0.099$	$0.075$	$0.083$	$0.169$	$+11.7$
H4 $\to$ H1	$0.075$	$-0.040$	$0.089$	$0.066$	$0.071$	$0.164$	$+12.6$
H2 $\to$ H5	$0.038$	$-0.018$	$0.053$	$0.042$	$0.045$	$0.084$	$-0.5$
H1 $\to$ H2	$0.030$	$-0.055$	$0.054$	$0.039$	$0.042$	$0.082$	$+9.8$

Appendix

Table 10: The nine swaps that a frozen comparison rejects and that have

T>0

, ordered by the observed gain. The ceiling gain

\Delta C

, the frozen return change,

T

, and the predicted gain

T-r_{2}

come from the post-trained checkpoints before adaptation. The gain is the adapted pair’s return minus the source pair’s return after 1,024 updates. Ranking is the adapted checkpoint’s return under the target minus its return under the source. Success is the change in task success from the source pair, in percentage points. The prediction for H0

\to

H1 was computed after its adaptation.

Calibration quantity	$m=64$ , $N=16$	$m=256$ , $N=1024$
Largest condition-level absolute bias in $M$	0.24805	0.00287
Largest family-level absolute bias in $M$	0.40320	0.00442
Worst-node return-difference RMSE	0.20259	0.00672
Mean actor queries per task and condition	356.9	42,553.3

Appendix

Table 11: Calibration of the measurement budget. Both columns use a pseudocount of

0.5

per action. Bias refers to the magnitude of the expected mismatch estimation error. Query counts include pilot actions and frequency samples at nonterminal states.

Figure 8: Measurement accuracy depends on the query budget and smoothing rule. Each point is a calibration candidate. (a) Largest absolute mismatch bias across deployment conditions. (b) Largest aggregate return-difference RMSE across adaptation nodes. Dashed lines mark calibration thresholds. The cross marks the initial budget and the star marks the selected estimator.

Estimator	Configuration RMSE	Recovery RMSE	Ranking RMSE
Ceiling minus measured mismatch	0.00230	0.00245	0.00286
Ceiling alone	0.23470	0.05744	0.05342
Frequency rollout, same query budget	0.00240	0.02672	0.02667

Appendix

Table 12: Prediction error on the fixed test tasks. Configuration RMSE averages over 72 checkpoint and configuration comparisons. Recovery and ranking RMSE average over seven saved adaptation checkpoints. No estimator reads exact action probabilities.

Figure 9: Measured mismatch predicts the return realized under a harness revision. Each point compares a candidate with its baseline on the fixed test tasks. The diagonal marks agreement between predicted and observed return changes. Circles and squares denote the H0-trained and H1-adapted checkpoints. (a) Ceiling minus measured mismatch. (b) Direct rollouts with action probabilities estimated from frequencies at the same actor-query budget. (c) Ceiling alone.

Figure 10: Ceiling and mismatch of five coding-agent harnesses for a frozen Qwen3.5-9B. Estimates of Table 6 on three Terminal-Bench Lite probe tasks with

\tau=0.1

. Each harness sits at its accuracy over all 39 tasks. The open circle marks

\widetilde{C}

, the filled circle marks

\widetilde{C}-\widetilde{M}

, and the bar between them has length

\widetilde{M}

Harness	Decisions	$\widetilde{C}_{h}$	$\widetilde{M}_{h}$	$\widetilde{C}_{h}-\widetilde{M}_{h}$
Terminus-2	4.9	0.431	0.134	0.297
mini-SWE-agent	8.6	0.631	0.335	0.296
DeepSeek Harness	8.6	0.590	0.382	0.208
Pi coding-agent	6.6	0.383	0.206	0.177
DeepAgents	14.2	0.261	0.429	$-0.168$

Appendix

Table 13: Terminal-Bench Lite probe results at

\tau=0.1

. Decisions are the mean number of model calls per complete target trajectory.

Harness	States	Inspect	Modify	Plan	Stop	Opaque	Invalid
Terminus-2	44	37.7 / 30.3	15.8 / 20.8	0.0 / 0.0	38.3 / 35.1	8.2 / 6.4	0.0 / 0.1
mini-SWE-agent	77	59.7 / 43.5	14.1 / 11.3	0.0 / 0.0	11.7 / 22.4	14.1 / 18.9	0.1 / 0.0
DeepSeek Harness	69	33.3 / 20.0	12.1 / 15.0	15.1 / 26.5	11.2 / 14.8	27.7 / 18.6	0.1 / 2.7
Pi coding-agent	46	48.3 / 37.0	20.4 / 22.4	0.0 / 0.0	12.9 / 21.0	18.0 / 16.6	0.0 / 0.4
DeepAgents	85	27.9 / 24.8	11.8 / 16.8	11.1 / 11.4	3.9 / 8.4	38.2 / 32.4	3.5 / 0.0

Appendix

Table 14: Action category shares (%) at the states of complete target trajectories on the Terminal-Bench Lite probe tasks, target / teacher.

Figure 11: Harness scores compared with task success. Each point is one harness evaluated with the same frozen Qwen3.5-9B on a shared task set. The panels compare

C

-M

, and

C-M

with rollout success. Dashed lines show linear fits.

LLM Agent Harnesses

LLM agent harnesses are the software and runtime systems that organize an agent’s tools, context, state, execution, and safeguards.

See all 389 papers

Oct 1, 2026

Finding the Right Fit: Model-Harness Interactions across Agent Tasks

Choosing an agent system means choosing both a language model and the harness through which it acts. We ask whether a strong model, harness, or pairing stays strong when the setting changes. We evaluate 66 configurations: four configurable harnesses (OpenHands, DeepSeek Harness, PI, and openJiuwen) paired with five models on TUA-Bench, ALE-CLI, and Terminal-Bench 4, plus the native Codex-GPT and Claude Code-Claude pairings. Model rankings reverse across harnesses. On Terminal-Bench 4, Claude leads GPT by 7.94 points in OpenHands but trails it by 30.16 points in PI. For four of the five models, the best harness changes from one benchmark to another, yet some pairings hold: openJiuwen gives Kimi its highest score on all three benchmarks, by 5.61 to 11.11 points. A model's own vendor harness is not reliably its best, and higher cost does not reliably buy a higher score. On Terminal-Bench 4, GPT scores higher under PI than under DSH at less than a quarter of the cost per task. Matched trajectories suggest why fit varies. Models start almost all repairs themselves, so much depends on whether the harness hands failures back in a form the model can use. GPT does best with PI's lean scaffold, while Kimi, which often issues malformed tool calls, does best in openJiuwen. We argue that the model, the harness, and the task should be evaluated together, and we release the harness adapters, evaluation code, and all 6,204 scored trajectories at https://github.com/liyix/finding-the-right-fit and https://huggingface.co/datasets/yixuanli97/finding-the-right-fit.

Yixuan Li, Yiyun Zhou, Yao Long Teng +6

College of Computing and Data Science, Nanyang Technological University, Singapore

Tool-Augmented Language Model Agents LLM Agent Evaluation AI Agent Benchmarks LLM Agent Harnesses

Ask PDF

Oct 14, 2026

Harness Policy Distillation: Aligning Agent Policies with Runtime Semantics

Deploying a general-purpose model as an agent means running it inside a product-specific runtime harness that decides which actions are executable, how state advances, and when the loop may terminate. Ordinary trajectory supervision shows which action was taken but leaves the runtime semantics behind that action implicit. We introduce Harness Policy Distillation (HPD), which makes these runtime control semantics an explicit training target alongside actions to align the harness policy. HPD replays reference trajectories through the harness to recover the runtime semantics that are implicit in ordinary trajectory data, and represents them as control documents at each model-call boundary. These documents provide privileged information for a teacher to generate actions, while harness verification ensures that the resulting targets are executable under the runtime contract. We jointly train a shared model on the runtime semantics recovered from control-state privileged information and native history, thereby internalizing these semantics into the model policy. At deployment, the trained model uses the original harness interface without requiring the control document or any additional supplement. On TRAJECT-Bench with the Pi harness, Qwen3-8B trained with HPD reaches 22.98% autonomous task success and 49.23% runtime compliance, against 3.09% and 19.90% for the unadapted model. Harness-verified teacher actions substantially improve the native deployment policy, while joint control-state supervision provides further gains in task completion and runtime compliance. HPD turns executable runtime semantics into dense supervision for learning harness-compatible actions at each decision boundary.

Junze Zhu, Weihao Chen, Zhiyang Zhou +2

National Key Laboratory for Novel Software Technology, Nanjing University, China · School of Artificial Intelligence, Nanjing University, China

LLM Agent Training Policy Distillation LLM Agent Harnesses Knowledge Distillation

Ask PDF

Oct 3, 2026

What Does a Harness Buy? Tokens, Mostly

A coding agent is a language model wrapped in a harness: the system prompt, the tool set, and the context management that turn a chat model into something that can work inside a repository. Production harnesses ship releases daily, vendors advertise pass-rate gains from harness changes, and leaderboards mix harnesses freely. What is rarely measured is how much the harness itself moves the score when the model is held fixed. We run five models through three production harnesses, Claude Code, mini-SWE-agent, and OpenCode, on SWE-bench Verified, and rerun the same configurations to calibrate how much a score moves when nothing changes but the run. On 447 tasks and the two models we ran there, Claude Code and mini-SWE-agent, the heaviest and the lightest harness, are equivalent within five points. On a 45-task hard subset and five models, swapping the harness flips as many tasks as rerunning the same harness, 13% in both cases, and the tasks a harness wins in one run are not the tasks it wins in the next. The one harness effect that clears the noise is a loss, not a gain: OpenCode trails by up to 9 points on the large pool, and on one model half of that gap sits in runs its output cap cut short. What the harness does decide is the bill. With the same model, the same tasks, and one price list, cost per task differs by up to 3x across harnesses. The gap is set at the first call, by the preamble of system prompt and tool schemas each harness sends with every step, and scaled by the number of steps; per-step growth and per-call tool output differ far less. The provider's price for cached input scales the bill and does not reorder it. The rerun data also give the resolution a harness comparison needs: at the discordance we observe, 45 tasks catch a 13-point gap only half the time and no gap with 80% power, and 447 tasks resolve 5 points, still coarser than the gains many harness changes claim.

Yangze Liu, Zhongyi Han

Shandong University

LLM Agent Efficiency LLM Agent Harnesses AI Coding Agents LLM Agent Evaluation LLM Agents Coding Agent Benchmarks

Ask PDF

Abstract

LLM Agent Harnesses

Explore similar work

Finding the Right Fit: Model-Harness Interactions across Agent Tasks

Harness Policy Distillation: Aligning Agent Policies with Runtime Semantics

What Does a Harness Buy? Tokens, Mostly