JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs

1 Introduction

• A small decision model can handle routine multiple-choice choices in a fraction of the time and price, while the LLM handles choices it is unsure about.Source: Figure 1 · p. 1
• Most harnesses call a large model at every step, making long tasks slow and costly despite many calls ending in multiple-choice decisions.Evaluating agents on one public leaderboard cost about $40,000.Source: §1 · p. 1
• The introduction names AgentDecisionBench and JevAgentHarness as the paper’s focus.Source: §1 · p. 2

8 decision points

• Fit a threshold per decision point.Source: §8 · p. 2

5 integration modes

• The process records JEV and answer-key labels for visited states, then tests whether each delegation is good and cheaper.Source: §5 · p. 2

3 benchmark levels

• The pipeline tests each kind of choice alone, then tests passing delegations together on new tasks.Source: Figure 2 · p. 2
• Decision models return probabilities over listed options from the task, history, and a multiple-choice question.Source: §3 · p. 2
• Fixed-question tests miss states following agent errors, where mistakes carry over and each choice changes what comes next.Source: §3 · p. 2
• AgentDecisionBench tests isolated questions, recorded agent steps without executing answers, and choices controlling a running agent.Source: §3 · p. 2
• On tasks unlike its training data, agents solved 91.5% of tasks with AgentDecision-4B tool choices and 82.0% with JEV.Source: §3 · p. 2
• Figure 3 maps eight agent-loop decision points across five integration modes, distinguishing choices of agent action from loop control.Source: Figure 3 · p. 3
• 81.1–89.0% cheaper and up to 2.5 times faster, customer-service and household agents using trained 4B models without losing 3 or more points of success.Source: §3 · pp. 2–3

2 Where and How to Delegate

• Eight recurring decision points cover tool, argument, and next-action choices, plus memory, reflection, stopping, LLM routing, and stored-note retrieval.A decision point is a kind of recurring choice whose options the harness can list from the state.Source: §2 · p. 3
• A decision model assigns probabilities to listed options, chooses the likeliest, and uses that option’s probability as its confidence.Open models can score all options in one forward pass.Source: §2 · p. 3
• The cascade uses a decision model only when confidence clears a threshold; otherwise the LLM decides, and the goal is the cheapest delegation set within 3 points of LLM-only success.Other modes let the model always decide, shortlist options, approve or reject the LLM, or trigger an optional step.Source: §2 · p. 3

2.1 Deciding what to delegate

• Delegation is worthwhile when confidence is at least q_L − ΔC/V, balancing LLM accuracy, saved-call price, and the cost of a wrong choice.The decision model may be less accurate than the LLM if the savings cover the extra risk.Source: §2.1 · p. 3
• With JEV and Flash, fitted thresholds were 1.00 for tools and 0.99 for arguments, so only near-certain answers were used.Thresholds were estimated per decision point from separate shadow-mode trajectories and set to the lowest confidence at which accepting was worthwhile.Source: §2.1 · p. 4

• AgentDecisionBench tests isolated questions, recorded steps, and choices carried out by a running agent, while tracking call price and time.The levels progress from fixed questions to choices that change what the agent encounters next.Source: §3 · p. 4
• At the isolated-question level, models answer multiple-choice conversions of BFCL v4 tool and argument items and When2Call items.The supplied passages specify 1,114 tool-choice items, 1,949 argument items, and 928 When2Call items.Source: §3 · p. 4
• At step level, models answer 5,314 shared states from 200 new DecisionGym tasks, but their answers are scored without changing the recorded runs.Answers are checked against acceptable next steps, so a mistake cannot affect the next state.Source: §3 · p. 4
• At task level, models make every tool choice in a running agent, so wrong choices alter later states and task success measures the outcome.The evaluation covers τ2-bench, ALFWorld, DecisionGym, SWE-bench Verified, Terminal-Bench 2.0, and Terminal-Bench-Science.Source: §3 · p. 4
• DecisionGym derives acceptable next steps at every solvable reachable state, including after detours, and separates training, threshold-fitting, and test seeds.Its scripted customer-service tasks include near-duplicate distractor tools and states where asking the customer is required.Source: §3 · p. 4

• The harness turns each listed choice into a multiple-choice question and uses the model’s answer only when confidence is high enough.Below threshold, the LLM decides as usual.Source: §4 · p. 4
• Tool calls are split into a tool choice and choices for fixed-list, yes/no, or context-copied arguments, leaving free-text arguments to the LLM.Any model that accepts Jev’s input format can be used.Source: §4 · pp. 4–5
• Thresholds are fitted from shadow-mode trajectories for each decision point and refitted when the LLM, decision model, or task rules change.An uncertain answer costs a cheap model call but does not risk the task because the LLM decides below threshold.Source: §4 · p. 5
• Adapters connect through existing harness hooks, and if an adapter or decision model fails, the harness runs as before.The integrations include routing, guards, triggers, and command hooks.Source: §4 · p. 5

• The training uses Qwen3-4B with LoRA because API decision models are billed per call and cannot learn harness rules, while fixed-question data rarely covers states after agent mistakes.Training proceeds through supervised fine-tuning followed by reinforcement learning in the agent loop.Source: §5 · p. 5
• 15,694 decision states from 587 training tasks include 30.9% states following a mistake, collected with a scripted policy that repeats development errors.The data collection aims to include states the agent reaches after errors.Source: §5 · p. 5
• When JEV ranks the correct answer, the target mixes 0.3 gold one-hot and 0.7 JEV probabilities; otherwise it uses the gold one-hot label alone.SFT fits the target with cross-entropy and a squared-error Brier term to keep probabilities calibrated.Source: §5 · p. 5
• GRPO trains the SFT model on its own task outcomes, rewarding success while reducing reward for LLM calls and irreversible wrong calls.Episodes run in groups of eight; a solved task with ten or more LLM calls earns half, and the fifteenth iteration is retained.Source: §5 · p. 5
• For evaluation, Stage 2 replaces JEV with the frozen AgentDecision-4B model and refits the thresholds.Source: §5 · p. 5

6 Experimental Protocol

• Stage 1 tests each delegation alone against the LLM on the same tasks, retaining only cheaper delegations that meet the non-inferiority test.A one-sided 5% test requires a task-success loss of less than 3 points, or 5 on τ 2-bench.Source: §6 · p. 5
• Eleven registered tests use Holm correction to keep the chance of any false pass at 5%.Models, tasks, and tests were preregistered, and the Stage-2 plan was registered before Stage-1 results were read.Source: §6 · p. 5
• Stage 2 selects the highest-saving Stage-1 delegation at each decision point, or keeps the LLM if none passed, and tests the combination on new tasks.The test set includes 500 new DECISIONGYM tasks and τ 2-bench telecom tasks.Source: §6 · p. 6
• Four Holm-corrected Stage-2 tests check combination performance and whether AgentDecision-4B can replace JEV.They test non-inferiority to the LLM alone and the best single delegation, plus comparisons with Intern-Decision-4B and JEV.Source: §6 · p. 6

7 Results

• 99.4% of DECISIONGYM tasks were solved by the LLM alone, so Table 1 reports harder settings.Table 1, Figure 4, and shifted tasks came from experiments added after Stage 1.Source: §7 · p. 6

7.1 RQ1: Do isolated scores predict success?

• 91.5% of tasks were solved by AgentDecision-4B, versus 69.5% for the best decision model on isolated questions.The isolated-question leader was less accurate on BFCL tool choice than AgentDecision-4B, while step-level accuracy ranked models nearly as task success did (Kendall correlation 0.84).Source: §7.1 · p. 6

7.2 RQ2: Which decisions can be delegated?

• A thresholded cascade saved cost at both tool and argument choices, whereas modes without a threshold lost success.The supplied passage does not give the savings or success changes for this comparison.Source: §7.2 · p. 6

AGENTDECISION-4B

• The fitted threshold cut cost by 14.2% without changing outcomes on any of 500 tasks and passed its registered test; a fixed threshold lost 21.5 points.With Flash-Lite, the fitted threshold left every tool choice to the LLM, so the cascade saved nothing.Source: AGENTDECISION-4B · p. 6

7.3 RQ3: What changes in deployment?

• 86.3% cheaper and 2.5 times faster, the DECISIONGYM agent with AgentDecision-4B did not lose 3 or more points of success.On ALFWorld, a 4B model trained there made the agent 89.0% cheaper and about twice as fast, with higher success; Table 1 summarizes harder settings.Source: §7.3 · p. 7, Table 1 · p. 7

7.4 RQ4: Can a 4B model learn the decisions?

• 96.0% of shifted tasks were solved by Qwen3-4B after training on agent states, up from 13.5%; StartLux-4B rose from 49.0% to 98.0%.SFT raised shifted-state step accuracy from 52.9% to 98.4%; the JEV combination failed non-inferiority, while replacing JEV with AgentDecision-4B passed at 81.1% lower cost.Source: §7.4 · pp. 7–8

7.5 Where delegation fails

• On hard tasks with 10 customers, AgentDecision-4B solved 3.0% of tasks against 88.0% for the LLM alone when facing a rule absent from training.Refitting the threshold restored success with JEV, and retraining AgentDecision-4B on the rule raised held-out accuracy from 76.5% to 95.3%; all 34 JEV failed tasks began with an irreversible wrong tool choice.Source: §7.5 · p. 8

8 Related Work

• JEV and open decision models score listed options, while REFLEX defers to the LLM below one global confidence threshold; this work fits a threshold per decision point.DecisionFlow structures LLM decisions.Source: §8 · p. 8
• AGENTDECISIONBENCH adds step- and task-level tests to decision benchmarks that score fixed questions and agent benchmarks that score whole episodes.The related work notes that models with equal benchmark scores can still make different decisions (Zhang et al., 2026).Source: §8 · p. 8
• Unlike routers that select a model for a call and agentic RL that trains the LLM, this work trains a decision model that scores listed options and defers when unsure.The related work connects this approach to learning to defer.Source: §8 · p. 8

9 Conclusion

• 81.1–89.0% cheaper and up to 2.5 times faster, agents using models trained on their own states lost fewer than 3 points of task success.These results support the paper’s hypothesis that decision models help when tested, used, and trained inside the agent.Source: §9 · p. 8
• Trained on agent states, AgentDecision-4B beat JEV inside the agent, while JevAgentHarness fitted thresholds on agent trajectories so the model handled most routine choices.Source: §9 · p. 8
• The paper’s results support evaluating decision models within running agents rather than relying only on isolated tests.Source: §9 · p. 8

Limitations

• The evaluation uses English tasks and mostly Gemini agent LLMs, with DeepSeek and OpenAI models in a few runs.Other languages or agent LLMs may make different choices worth delegating.Source: Limitations · p. 9
• The main tests establish success drops below a fixed margin—3 points, or 5 on τ 2-bench—not equal success.Equal success would require an equivalence test.Source: Limitations · p. 9
• A new agent, new tools, or a new task family calls for refitting thresholds; best accuracy also calls for training on that agent’s states.The authors did this for ALFWorld.Source: Limitations · p. 9
• In coding agents with free-text actions, the decision model can only choose which LLM writes a step, and savings were 7.0–10.5%.Routing file-reading steps to a fast LLM gave no reliable saving because the decision model could hardly distinguish them from other steps.Source: Limitations · p. 9
• The study did not compare with learned routers such as RouteLLM on the same tasks, and recorded prices and latencies vary with providers, hardware, and serving.GRPO added no measurable gain over SFT in the authors’ runs.Source: Limitations · p. 9

Ethical Considerations

• Adapters control cost and speed, not security; adversarial observations and policy injection were untested, and cheaper control may lower the cost of misuse at scale.With the cheaper Flash-Lite agent LLM, the stop guard accepted 40.7% of the LLM’s own wrong attempts to end the task, so it should not be the only control on irreversible actions. Results therefore apply to benign states only (Xu et al., 2026; Hu et al., 2026; Wu and Lim, 2026a).Source: Ethical Considerations · p. 9

Compute, Data and AI Use

• DECISIONGYM tasks use generated customer data, public benchmark items, and existing answer keys or checkers; no human annotators or study participants took part.The work used cloud-hosted open models on one NVIDIA L40S, A10G, or L4 GPU, and AI assistants helped with code, references, and writing under the authors’ direction; the authors take responsibility for the content.Source: Compute, Data and AI Use · p. 15

A Delegation Theory

• Fitted thresholds beat the default τ = 0.8 in 70 of 72 cells, but temperature scaling returned 0 of 3 over-confident cells to the cost–success front.In a registered counterexample to H3, JEV’s Flash tool cascade delegated irreversible, copy-dependent, bounded-recoverable decisions while remaining non-inferior (∆= +0.0 points) at 14.2% savings.Source: A Delegation Theory · p. 17

B.1 Suites and the DECISIONGYM Environment

• DECISIONGYM’s generator provides an answer key from any solvable state, uses disjoint seed ranges, varies task difficulty, and uses scripted customers to limit episode costs to model calls.Evaluation uses the two hardest task families, clarify and multi, with strict-mode failures for state-changing calls that deviate from gold calls or precede a required question.Source: §B.1 · p. 17

B.2 Example Tasks

• The example tasks show that the answer key accepts progress-making steps using only values the agent has seen, while detours are allowed but cost a step.Source: §B.2 · pp. 17–18, Table 6 · p. 19

B.3 Typed Questions and Decider Prompts

• The LLM decider estimates option probabilities by summing log-probabilities of option-label tokens and renormalizing; trained models use a full-logit readout.When log-probabilities are unavailable, the generated token receives all probability mass, while verbalized confidence is assigned to the answer and the remainder is spread uniformly.Source: §B.3 · p. 18

B.4 Prompts

• Harness branches become typed questions over the same state text; decision models fill bounded or copied arguments, while free-text arguments go to the LLM.The prompts contain no label for the decision being asked, and training rows use the trained-model format without adding a separate prompt.Source: §B.4 · pp. 18, 20, Figure 7 · p. 19
• JEV, OpenJev, and StartLux-Decision build prompts on their own servers, while other models use supplied prompts; Intern-Decision handles k questions in one forward pass.For Intern-Decision, the vendor’s licensed prompt compiler encodes state as JSON and reads each answer at its decision placeholder.Source: §B.4 · p. 20
• DecisionGym prompt examples cover cancel, refund, stock, shipping, clarification, and multi-request tasks.The multi-request prompt asks the agent to process requests in order and report the total amount processed.Source: §B.4 · p. 21
• Mini-SWE-agent reflection, action-type, and stop-trigger hooks were implemented but not run; its SWE-bench prompts remained unchanged.The prompts followed mini-SWE-agent’s SWE-bench configuration (Yang et al., 2024).Source: §B.4 · p. 23

B.5 Gold Labels, Probes and Measurement Arms

• Gold labels classify actions by whether they advance, derail, prematurely end, or correctly complete a task; measurement arms run gates in shadow and probes are scored but never executed.A gate is estimable for H2 only with at least 30 positive and 30 negative labels and both live and matched-rate random arms; the reflection-trigger test has no headroom when its one-step ceiling’s 95% CI lies within ±3.3 points.Source: §B.5 · p. 24

B.6 Benchmark Difficulty: Hard States

• Off-the-shelf decision models selected acceptable tools in 48.3–93.9% of common states and 42.7–82.7% of hard states, while Flash alone solved 99.4% of 500 fresh Stage-2 tasks.The hard subset was defined post hoc as states after an off-plan step or where the next tool was ask_user; generator-trained models were near ceiling on every subset.Source: §B.6 · p. 24, Table 8 · p. 25

C JEVAGENTHARNESS

• JEVAgentHarness configures a mode, threshold, and decision-model backend per decision point; under CASCADE, the LLM supplies the full call below the tool threshold and fills remaining arguments otherwise.The training readout obtains option probabilities from prefill logits, while the harness records configuration hashes and replays cached decision-model cost and latency. Prompts also specify guards for tool choices, command safety, verified submission, stopping, and reflection.Source: §B.4 · p. 23, C JEVAGENTHARNESS · p. 25, Figure 8 · p. 25

C.2 JEVAGENTHARNESS in mini-SWE-agent (Exploratory)

• 229 versus 232 tasks were resolved by JEV-routed versus strong-LLM-only mini-SWE-agent, at 10.5% lower cost, but non-inferiority was not established.The additional 200 instances saved 4.8%; with Gemini 3.8 Flash in both roles, the fast LLM alone resolved more tasks than either other arm. In separate exploratory final-harness runs, decision-model queueing took 66.1% of episode time when four episodes shared a GPU.Source: §C.2 · p. 26, §C.1 · p. 26, Table 10 · p. 26

C.3 Routing in Two More Open-Source Harnesses

• In pi and SWE-agent, routing saved little cost and no time; it kept success in pi, while routed SWE-agent used 33.4 versus 29.3 steps per instance.A proxy pilot routed 0.7% of mini-SWE-agent requests without changing success; JEV’s read-only predictions poorly separated read-only from other steps (AUROC 0.56 in pi and 0.70 in mini-SWE-agent), so none of the pilots qualified for a full run.Source: §C.3 · p. 26, Table 12 · p. 26

C.4 Roles of a Decision Model in a Harness

• At 582 states where asking was correct, AgentDecision-4B V1 chose correctly in 100.0%, JEV in 64.4%, and Intern-Decision-4B in 59.6%; JEV’s stop guard rejected 52 of 85 wrong first attempts, but only 4 episodes then succeeded.On 1,000 Flash calibration decisions, confidence-gate AUROC was 0.74 for tool choice and 0.89 for arguments; a Flash check cost 0.58 and 0.47 times the respective saved-call price, making net savings negative. A second decision-model question raised savings only for arguments.Source: §C.4 · p. 27

D.2 The Training Chain

• On-policy distillation was dropped after lowering development and recovery-state accuracy; the retained SFT v1 plus GRPO checkpoint passed development gates, but no gain from GRPO was isolated.SFT v1 already matched AgentDecision-4B V1 on common-state accuracy, task-level success (100.0%), and temperature-scaled ECE on the saturated development slice.Source: §D.2 · p. 28

D.3 Recipe Takeaways by Stage

• SFT on the agent’s own states gave almost all the gain for about two GPU-hours, while GRPO added no measurable gain; public tool-use data raised isolated accuracy but lowered success on shifted tasks.Adding post-off-plan states (30.9% of data v1) raised development accuracy on those states from 92.2% to 99.8%. A second track’s SFT v2 succeeded on 100.0% of 500 fresh tasks against 99.4% for LLM-only, at 54.8% lower cost including its GPU; OPD and GRPO lowered success on a harder 200-task development slice.Source: §D.3 · pp. 28–29, Table 15 · p. 29

D.4 Which Parts of the Recipe Matter

• GRPO without SFT made 98.1% of training decisions itself yet remained at the untrained model’s level on every measure, while evidence about which states to train on was mixed.On common states it scored 48.5% against 50.1% before and 99.2% after SFT; 75.0% of task groups had no episode-level reward signal.Source: §D.4 · p. 29

D.6 DECISIONGYM-Hard

• 95.3% held-out accuracy on cancel-or-ship states followed one epoch of tier-rule training, up from 76.5%, while common-state accuracy stayed at 99.1% against 99.2%.The trained model accepted no wrong cancel-or-ship decision in the agent, but the paired test did not establish non-inferiority at the 3-point margin.Source: §D.6 · p. 29

D.7 The Training Recipe on ALFWorld

• 95.0% of 140 valid_seen ALFWorld games were solved versus 90.7% for Flash alone, at 89.0% lower cost; on valid_unseen, success was 89.6% versus 88.8% and non-inferiority was not shown.The unchanged cascade took every action in valid_seen and 99.4% in valid_unseen; valid_seen shares rooms with training, while valid_unseen covers four unseen rooms.Source: §D.7 · pp. 29–30

D.5 A Sub-1B Decision Model with a Scoring Head

• 100.0% of 200 fresh DECISIONGYM tasks were solved by the 0.8B scoring-head model, matching AGENTDECISION-4B V1 at 64.2% lower cost than Flash alone.Arm A matched AGENTDECISION-4B V1 on common states, but its fitted thresholds were within 10^-9 of 1; Unsloth’s decision-model code is AGPL-3.0 and was not redistributed.Source: §D.5 · p. 29, Table 18 · p. 30

D.8 Agent-State Accuracy on Public Benchmarks

• 64.7% agent-state accuracy for AGENTDECISION-4B versus 61.1% for its untrained base did not transfer DECISIONGYM-trained models to ALFWorld; model rankings also differed on τ 2-bench.The ALFWorld-trained model led even in unseen rooms, while the τ 2-bench ranking had Kendall rank correlation −0.43 over 8 models.Source: §D.8 · p. 31

D.9 More Training on Public Tool-Use Data

• 100.0% of 200 fresh tasks were solved by both AGENTDECISION-4B and StartLux-Decision-4B after public-data training, with common-state accuracy non-inferior to AGENTDECISION-4B V1.AGENTDECISION-4B reached 85.1% isolated accuracy; StartLux-Decision-4B reached 85.5%, +2.3 points above its base.Source: §D.9 · p. 31

D.10 Train/Evaluation Hygiene

• The training/evaluation audit found no shared training and evaluation seeds or repeated task worlds or gold plans, but disclosed possible exploratory exposure and unexcludable data overlap.3.3% of common-state occurrences repeated a training state verbatim, with AGENTDECISION-4B V1 scoring 99.2% both with and without them; closed training data for Intern-Decision and StartLux-Decision prevent ruling out overlap with public L1 items.Source: §D.10 · pp. 31–32
• On DECISIONGYM common states, template lookups could get almost every state right, while AGENTDECISION-4B V1 scored 98.3% on shifted states whose correct tool was renamed.Only the argument shortlist and action cascade changed agent choices at lower cost among the reported exploratory cells.Source: §D.10 · p. 32, Figure 10 · p. 34, §G.2 · p. 34

G.5 Notes on Figure 4 and Tables 1–3

• Stage-2 rows report estimates, 95% intervals, null boundaries, and one-sided p-values; StartLux-4B costs $0.00217 per episode and saves −1.3% at Jev’s API price or 3.8% at zero GPU price.The V1 GPU price is $0.00196 per episode.Source: Table 27 · p. 36
• Secondary Stage-2 tests report success differences, the share of tool decisions delegated, and model-call seconds per episode, with negative seconds indicating slower performance.The table also identifies alternate episode parallelism and reference conditions for selected rows.Source: Table 28 · p. 37
• Table 1 paired tests define wins and losses as tasks solved only by the delegated arm or reference, respectively, and assess non-inferiority using a 3-point margin (5 points on τ 2-bench).Non-inferiority is indicated when the one-sided p-value is ≤0.05; the “100 + 200 new” results pool a routing run with its extension.Source: Table 30 · p. 37

G.6 Evaluation Levels and Reach

• Across five pre-specified models, isolated and task-level rankings disagreed (Kendall rank correlation −0.200).For ten off-the-shelf models, isolated tool-choice and common-state accuracy rankings correlated with task success at 0.18 and 0.54, respectively.Source: §G.6 · p. 37, §G.7 · p. 38, Table 31 · p. 38

G.7 More Decision Models as Baselines

• Task failures often ended at a first irreversible wrong choice; must-ask accuracy was only 26.0–27.0% for three models, despite must-ask states comprising 5.0% of the corpus.Counting first irreversible errors improved ranking correlation over common-state accuracy, but still missed task-level success by 14.0 points on average.Source: §G.7 · pp. 38–39, Table 34 · p. 39

G.8 Variants of the Decision Model

• Majority voting, confidence-weighted voting, and agreement gates did not reach AGENTDECISION-4B V1; public-data continuation also left the isolated gap and When2Call as weaknesses.Both continued-training arms were non-inferior to V1 on common states at the registered 1-point margin.Source: §G.8 · p. 39, Table 35 · p. 39, Table 36 · p. 39
• At equal accuracy, the scoring head was 2.7 times faster, used 2.5 rather than 4.5 GB of memory, and accepted no wrong common-state tool choices at its threshold.The one-token read-out threshold accepted 66 wrong tool choices on common states.Source: §G.8 · p. 40, Table 37 · p. 40
• 100.0% of the 200 final-harness tasks were solved with the deferral-aware model, versus 99.5% with Flash alone, at 89.3% lower cost.The model took every tool choice; the registered criterion of higher coverage and success within 3 points held.Source: §G.8 · p. 40

G.9 A Shifted DecisionGym Split

• Under shift, generator-trained 4B models lost little and remained ahead of off-the-shelf models at both evaluation levels, while in-domain thresholds did not establish harness non-inferiority.The 4B models stayed confident and took most tool choices; only the 0.8B head’s learned deferral handed choices back, and none of the three met non-inferiority.Source: §G.9 · p. 40, Table 39 · p. 40, Table 40 · p. 41

G.10 Hard Tasks (Exploratory)

• No arm reached the registered verdict on the exploratory SWE-bench Verified and Terminal-Bench-Science hard-task evaluation.The later step-routing runs superseded the routing row.Source: §G.10 · p. 41, Table 41 · p. 42

H Extended Related Work

• AGENTDECISIONBENCH measures decision-model performance on isolated questions, recorded states, and live tasks to expose differences that static parity can hide.The related-work discussion distinguishes this evaluation across decision points from prior reports of static parity in browser agents and legal documents.Source: H Extended Related Work · pp. 41–42