acceptodds
Under review as a conference paper at ICLR 2027

CollabOrchestra: Learning to Orchestrate Human-Agent Teams

Abstract

AI agents are increasingly deployed in occupational contexts, where many real-world scenarios involve humans and agents working together. Effective human-agent collaboration can outperform either party alone while preserving human agency, but requires understanding each other's capabilities and deciding when humans should delegate, verify, or intervene. We present **CollabOrchestra**, a practical method for learning an orchestrator policy that delegates work across a human-agent team. CollabOrchestra casts orchestration as *adaptive workflow plan generation*, where the orchestrator constructs and revises a sequence of steps specifying goals, assignees, and handoff deliverables. Our training proceeds in two stages: we first profile agent workers through sandboxed execution and learn a feasibility predictor from their execution traces, then train the orchestrator through multi-turn reinforcement learning to jointly optimize plan quality and execution feasibility. Across GDPval, JobBench and their underspecified variants, CollabOrchestra outperforms prompting-based orchestrators in both plan quality and execution outcome. The resulting team also outperforms the best solo agent, improving Task Performance from 0.474 to 0.744 on GDPval and from 0.260 to 0.286 on JobBench. Our analysis shows that CollabOrchestra learns to assign humans diverse, task-dependent roles, including directly performing work when appropriate. In a user study on participants' own tasks, users prefer CollabOrchestra-trained plans 16% more and report high collaboration legibility.

Primary area:
reinforcement learning
Keywords:
intelligent delegation, human-agent collaboration, multi-turn RL

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UYraOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CollabOrchestra, a framework for training language-model orchestrators to coordinate human-agent teams at the workflow level. The orchestrator constructs and revises structured plans specifying subtask goals, worker assignments, and so on. The method consists of two stages. First, the authors profile agent workers through sandboxed execution and train a lightweight feasibility predictor to estimate whether a particular worker can complete a proposed subtask. Second, they train the orchestrator using multi-turn GRPO with simulated users, optimizing a combination of LLM-judged plan quality and predicted execution feasibility.

Critical strengths
  1. I think the problem formulation is useful. The paper treats human–agent delegation at the workflow level, with the human as an assignable worker rather than a passive recipient. This is a practical abstraction for long-horizon occupational tasks.
  2. The training recipe is practical. They use an offline, execution-grounded feasibility predictor to replace costly rollouts during RL, and the paper checks that planning gains carry over to execution.
Critical weaknesses
  1. The bottleneck feasibility in Table 1 is evaluated using the same learned predictor that supplies the training reward. I don't think improvements in this metric are independent evidence of improved execution feasibility. 2.The “team” includes an extra gemini-3-flash agent playing the human. The team therefore gets more compute and more review passes than a single solo agent. I don't think the gain can be credited to orchestration. Although the paper provides ablations analyzing human-assigned roles, it does not sufficiently isolate the contribution of each reward component to downstream delivery rate and task performance.
Questions for authors
  1. Does the predicted bottleneck feasibility correlate with observed execution success on held-out workflow steps or complete workflows?
  2. Could the authors report delivery rate and task performance for the ablations without the feasibility reward and without the user-involvement component?
8 comments

We thank the reviewer for the constructive feedback. We are glad the reviewer finds that the problem formulation is useful and provides a practical abstraction for long-horizon occupational tasks. We also appreciate the reviewer's view that the training recipe is practical: the offline, execution-grounded feasibility predictor replaces costly rollouts during RL, and the paper checks that planning gains carry over to execution. We address each concern below.

The bottleneck feasibility in Table 1 is evaluated using the same learned predictor that supplies the training reward. I don't think improvements in this metric are independent evidence of improved execution feasibility.

Does the predicted bottleneck feasibility correlate with observed execution success on held-out workflow steps or complete workflows?

We check the correlation between Bottleneck Feasibility and delivery status (0 or 1) for every executed plan, pairing each orchestrator plan with its end-to-end execution. The point-biserial correlation r reveals that a delivered plan is scored significantly higher than an undelivered one.

DatasetPoint-biserial correlation
GDPval0.121 (p < 1e-6)
JobBench0.057 (p = 0.03)

We agree that the Bottleneck Feasibility numbers in Table 1 are not independent evidence of execution feasibility, since the policy was optimized against the same predictor. We use Table 1's Bottleneck Feasibility as a diagnostic of what the policy learned. The independent evidence comes from two other sources:

  1. The predictor's validity on held-out data which are grounded in actual execution results. In Appendix E, we validate the feasibility predictor on 19,173 agent-executed workflow steps from six agent frameworks across 12 benchmarks (Wang et al., 2026). Using workflow-disjoint 5-fold cross-validation, the encoder-based predictor reaches AUROC 0.74 / Brier 0.24, outperforming a zero-shot LLM (0.59 / 0.31) and an LLM post-trained with GRPO on the same task (0.59 / 0.26). So the predictor's scores carry real information about whether held-out steps succeed.
  2. Delivery Rate of actual execution outcomes. As shown in Table 2, when plans are actually executed with real agent workers, CollabOrchestra improves Delivery Rate over Rubric Prompting in all 12 settings, with the largest gains where predicted feasibility also improves most (e.g., GDPval-underspecified: 0.558 → 0.952 for Qwen3-30B). These outcomes are measured by execution, not by fϕf_\phi

The “team” includes an extra gemini-3-flash agent playing the human. The team therefore gets more compute and more review passes than a single solo agent. I don't think the gain can be credited to orchestration. Although the paper provides ablations analyzing human-assigned roles, it does not sufficiently isolate the contribution of each reward component to downstream delivery rate and task performance.

We added a multi-agent baseline and showed that multi-agent systems without hidden task description, cannot achieve desired task performance as if informed users are involved.

Experiment protocol. We implement the multi-agent baseline by using the same workflow plan but giving the "user" steps to an additional GeminiCLI agent. The additional agent receives the original task prompt, the full workflow plan, the current step's goal and deliverables, and the outcomes of previous steps. It does not receive the user profile or the user-only information for underspecified tasks. Each step has a 15-minute timeout instead of 3 minutes, so the baseline gets at least as much compute as the user simulator.

Results. Execution results of the simulated human-agent team and the multi-agent baseline using the same workflow plans from CollabOrchestra. pp-values are from permutation tests on Task Performance. Each is shown on the row of the higher-scoring setting. n.s. = not significant.

OrchestratorMethodGDPval Delivery RateGDPval Task Perf.GDPval-underspecified Delivery RateGDPval-underspecified Task Perf.
Qwen3-4B CollabOrchestraSimulated Human-Agent Collaboration0.9700.7410.9520.645 (p <1e-4)
Qwen3-4B CollabOrchestraMulti-Agent Baseline0.9640.742 (n.s.)0.9800.472
Qwen3-30B-A3B CollabOrchestraSimulated Human-Agent Collaboration0.9570.5970.9520.474
Qwen3-30B-A3B CollabOrchestraMulti-Agent Baseline1.000.786 (p = 0.0003)0.9850.481 (n.s.)
gpt-oss-120b CollabOrchestraSimulated Human-Agent Collaboration0.9260.7440.9300.703 (p <1e-4)
gpt-oss-120b CollabOrchestraMulti-Agent Baseline0.9740.753 (n.s.)0.9850.495

When tasks are fully specified, the multi-agent baseline is about as good or slightly better, largely by delivering more often. For underspecified tasks, however, replacing the informed user with an additional agent substantially reduces task performance for Qwen3-4B (0.645 vs. 0.472, pp < 1e-4) and gpt-oss-120b (0.703 vs. 0.495, pp < 1e-4), despite the additional execution budget.

These findings qualify our claims and address the compute confound: additional agent compute can reproduce the benefits of human-agent collaboration when tasks are fully specified, but it cannot reliably substitute for user-specific knowledge and preferences that are absent from the initial task description. Since real-world tasks often depend on such information, effective orchestration must go beyond allocating more agent compute to meaningfully involve informed users when their contributions are necessary.

We execute the workflow plans from two ablations from Section 5.2:

  • “w/o Feasibility Reward”, which removes the feasibility term from the training reward;
  • “w/o User Involv. Rubric”, which removes the Use Involvement Appropriateness component for the Plan Quality reward

The results are as follows:

Orchestrator Base ModelMethodGDPval Delivery RateGDPval Task Perf.GDPval-underspecified Delivery RateGDPval-underspecified Task Perf.
gpt-oss-120bCollabOrchestra0.9260.7440.9300.703
gpt-oss-120bCollabOrchestra w/o feasibility reward0.6940.5560.7300.507
gpt-oss-120bCollabOrchestra w/o user-involvement reward0.9590.7420.9230.651

Feasibility reward. Removing the feasibility term causes a large drop in Delivery Rate on both benchmarks (GDPval: 0.926 → 0.694; GDPval-underspecified: 0.930 → 0.730). Task Performance drops with it (0.744 → 0.556; 0.703 → 0.507). This confirms that the feasibility reward is critical for producing plans that can actually be executed.

User-involvement rubric. Removing the User Involvement Appropriateness effects Task Performance on GDPval-underspecified where the simulated user holds information the orchestrator needs. In this setup performance drops from 0.703 to 0.651, because without the rubric the orchestrator is less likely to involve the user at the points where that information is needed.

Reviewer_oCYKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CollabOrchestra: learning an LLM-based orchestration policy that coordinates human and AI workers at the workflow level. The orchestrator creates and revises structured plans consisting of goals, assigned workers, and expected deliverables, and interacts with users to clarify requirements and coordinate execution. The proposed training method has two stages. First, the authors profile agent capabilities using sandboxed executions and train an offline feasibility predictor to estimate whether individual agent-assigned steps can succeed. Second, they train the orchestrator using multi-turn GRPO with simulated users, optimizing a composite reward combining LLM-judged plan quality and predicted execution feasibility. They conducted experiments on GDPval, JobBench, and underspecified variants of both datasets, across three orchestrator backbones.

Critical strengths
  • The paper tackles the important problem of deciding how humans and AI agents should share work on long-horizon tasks. Workflow-level delegation between humans and agents is timely. Representing orchestration as adaptive workflow planning, with explicit worker assignments and deliverables, is a reasonable abstraction for coordinating heterogeneous workers. The combination of execution-grounded capability profiling and multi-turn RL is technically well motivated and avoids expensive end-to-end workflow execution during every training rollout. This is likely to be useful to others building orchestrators.

  • the authors present strong empirical coverage with evidence beyond plan-quality judgments. The evaluation covers three orchestrator backbones, two prominent benchmarks and their underspecified variants, and both planning and execution metrics. The improvements are particularly substantial for delivery rate on underspecified tasks (Table 2), where asking users for missing information is naturally important. The authors also include reward ablations, analyses of human roles, and a real-user study. Together, these experiments provide useful evidence that the learned policy changes delegation behavior and can improve execution under the simulated setting.

Critical weaknesses
  • The primary execution experiments do not establish benefits from collaboration with actual humans. In Section 4.1 and Appendix F, human-assigned steps are executed by a Gemini CLI agent powered by Gemini-3-Flash. Thus, the reported improvements over the best solo agent arise from a multi-worker system in which the human is itself simulated by another capable agent. This is useful for testing orchestration, but it is not sufficient to establish that an actual human-agent team would achieve comparable gains.
  • Task Performance is also computed only on delivered cases, so differences in Delivery Rate can bias it.
  • Results are mixed in places: for Qwen3-30B on GDPval, CollabOrchestra gets 0.597 vs. 0.659 for Rubric Prompting, and gpt-oss-120b on JobBench-underspecified is 0.214.
  • The real-user study provides limited evidence that the learned orchestration policy improves actual task outcomes or user agency. In Section 6, participants compare plans from both orchestrators, but execute only the CollabOrchestra plan. Consequently, the study cannot directly compare real-user task completion, collaboration burden, or agency under the two approaches. The preference difference is relatively small (31 versus 22 out of 55 sessions, with two ties), and the paper does not establish that this difference is statistically reliable.
  • Furthermore, the positive questionnaire ratings are obtained only after interacting with CollabOrchestra, without a comparison condition. The findings therefore provide encouraging initial usability evidence, but not a controlled demonstration of superior real-world human-agent collaboration.
Questions for authors
  • How do the Table 2 Task Performance comparisons change if undelivered tasks are scored as 0, i.e., overall rather than conditional on delivery?
  • Can you add a multi-agent baseline with matched compute and step budget, e.g., the same plan but with the "user" steps given to a third agent without the hidden task description, or a solo agent with the same total budget?
  • Can you report variance across seeds or bootstrap intervals for the Table 2 execution results?
  • In the user study, 31 vs. 22 preferences out of 53 non-tied sessions doesn't look statistically significant. Did you test this? The Likert ratings were collected only for the CollabOrchestra condition, so how should they be read without a baseline condition?
5 comments

Thank you for recognizing our work as addressing a timely and important problem, providing a reasonable abstraction for coordinating human-agent teams in long-horizon tasks, and having strong empirical coverage in our evaluation.

We appreciate your view that combining “execution-grounded capability profiling with multi-turn RL is technically well motivated, avoids expensive end-to-end execution during training, and is likely to be useful to others building orchestrators”, which is the core of our method.

We address your questions below:

Necessity of simulated experiments

W1: The primary execution experiments do not establish benefits from collaboration with actual humans. … This is useful for testing orchestration, but it is not sufficient to establish that an actual human-agent team would achieve comparable gains.

We agree that the simulated execution experiments are not a substitute for real human teams, and we complement the simulation result with the real-user study, in which 24 participants each brought their own tasks (55 sessions) in Section 6.

However, simulation is necessary for the main comparison. Isolating the effect of the orchestration policy requires holding the task, the hidden information, and the human worker fixed across conditions Real participants cannot provide this, because the same person cannot carry out the same task under several plans without learning or fatigue effects. We constrain the simulated human so that it behaves like a bounded collaborator rather than a free extra agent:

  1. It acts only on steps explicitly assigned to it
  2. Each user step has a 3-minute timeout contrasting to 15 minutes for agent steps, to model bounded human effort;
  3. On underspecified tasks, the missing requirements are available only to the simulated user, so the orchestrator gets them only by involving the user well.

Task Performance

W2: Task Performance is also computed only on delivered cases, so differences in Delivery Rate can bias it.

Q1: How do the Table 2 Task Performance comparisons change if undelivered tasks are scored as 0, i.e., overall rather than conditional on delivery?

We appreciate your question. Here is the result of reporting Task Performance on all tasks not conditioned on the delivered tasks.

ModelGDPval Delivery RateGDPval Task Perf. on AllGDPval-underspecified Delivery RateGDPval-underspecified Task Perf. on AllJobBench Delivery RateJobBench Task Perf. on AllJobBench-underspecified Delivery RateJobBench-underspecified Task Perf. on All
Solo Gemini CLI Agent (gemini-3-flash)0.940.446//0.8670.225//
Prompting Baseline (Qwen3-4B-Instruct-2507)0.9080.6430.650.2950.9610.2190.8520.287
Learning to Team Up (Qwen3-4B-Instruct-2507)0.970.7190.9520.6140.9610.2340.8440.297
Prompting Baseline (Qwen3-30B-A3B-Instruct-2507)0.8290.5460.5580.2240.9690.2340.6880.157
Learning to Team Up (Qwen3-30B-A3B-Instruct-2507)0.9570.5710.9520.4510.9840.2490.9060.333
Prompting Baseline (gpt-oss-120B)0.6370.3320.5120.1900.4770.0660.4610.065
Learning to Team Up (gpt-oss-120B)0.9260.6890.9310.6540.9690.2300.9220.271

We report the Task Performance conditioned on Delivery Rate because it isolates failure of completing the task (e.g., an infeasible plan leads to failure in completing the task for the human-agent team) and low output quality (e.g., the agent can still complete the final outcome without getting hidden information from the user but the task performance is very low).

W3: Results are mixed in places: for Qwen3-30B on GDPval, CollabOrchestra gets 0.597 vs. 0.659 for Rubric Prompting, and gpt-oss-120b on JobBench-underspecified is 0.214.

As shown in Table 2, on three base models and four datasets, CollabOrchestra outperforms Rubric Prompting on Delivery Rate on all setups and on Task Performance on 11 out of 12 setups, which supports the main conclusion of the work.

For JobBench-underspecified, it includes challenging work-relevant tasks and the task description only includes partial information. Under the same orchestrator base model, CollabOrchestra still outperforms Rubric Prompting on this dataset.

Q3: Can you report variance across seeds or bootstrap intervals for the Table 2 execution results?

Our results are stable across seeds and we report the 95% bootstrap confidence interval (N=10,000) below:

Orchestrator Base ModelModelGDPval Delivery RateGDPval Task Perf.GDPval-underspecified Delivery RateGDPval-underspecified Task Perf.JobBench Delivery RateJobBench Task Perf.JobBench-underspecified Delivery RateJobBench-underspecified Task Perf.
/Best Solo Agent0.94 [0.905, 0.974]0.474 [0.417, 0.535]//0.867 [0.805, 0.922]0.26 [0.213, 0.306]//
Qwen3-4BRubric Prompting0.908 [0.866, 0.945]0.708 [0.664, 0.751]0.65 [0.585, 0.714]0.454 [0.402, 0.507]0.961 [0.922, 0.992]0.228 [0.178, 0.278]0.852 [0.789, 0.906]0.337 [0.274, 0.401]
Qwen3-4BCollabOrchestra0.97 [0.939, 0.994]0.741 [0.697, 0.781]0.952 [0.915, 0.982]0.645 [0.598, 0.692]0.961 [0.922, 0.992]0.243 [0.192, 0.296]0.844 [0.781, 0.906]0.352 [0.289, 0.415]
Qwen3-30B-A3BRubric Prompting0.829 [0.779, 0.876]0.659 [0.608, 0.707]0.558 [0.488, 0.622]0.402 [0.348, 0.454]0.969 [0.938, 0.992]0.242 [0.196, 0.289]0.688 [0.609, 0.766]0.228 [0.169, 0.287]
Qwen3-30B-A3BCollabOrchestra0.957 [0.927, 0.988]0.597 [0.544, 0.651]0.952 [0.916, 0.982]0.474 [0.427, 0.526]0.984 [0.961, 1.000]0.253 [0.201, 0.307]0.906 [0.852, 0.953]0.367 [0.312, 0.422]
gpt-oss-120bRubric Prompting0.637 [0.567, 0.696]0.521 [0.459, 0.571]0.512 [0.447, 0.576]0.371 [0.316, 0.423]0.477 [0.391, 0.562]0.138 [0.096, 0.185]0.461 [0.375, 0.547]0.141 [0.096, 0.190]
gpt-oss-120bCollabOrchestra0.926 [0.889, 0.959]0.744 [0.703, 0.783]0.931 [0.894, 0.963]0.703 [0.663, 0.742]0.969 [0.938, 0.992]0.237 [0.186, 0.288]0.922 [0.875, 0.961]0.294 [0.244, 0.345]
Reviewer_m76nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper introduces CollabOrchestra, a method for training an orchestrator to allocate complex tasks to humans and AI agents.

Training proceeds in two stages: learning a feasibility predictor from agent execution records, followed by multi-turn reinforcement learning with simulated users using GRPO to jointly optimize plan quality and predicted execution feasibility.

Critical strengths
  • The paper addresses an important problem in human–agent collaboration: how to allocate work and determine when human participation is valuable.

  • The experimental evaluation is reasonably comprehensive, assessing both plan quality and execution outcomes, including delivery rates and final artifact quality. A real-user study with participant-provided tasks extends these evaluations with user experience.

Critical weaknesses
  • The human worker in the main execution experiments is simulated by GeminiCLIAgent. The share of human-assigned steps increases from 15.6% to 32.9% in the user study without a corresponding comparison of "human" time or workload.
  • The plan-quality rubric was iteratively refined using only 20 human-annotated comparisons, with no reported held-out human validation. Agreement across LLM judges does not mean alignment with human preferences, leaving the reliability of this training reward insufficiently validated.
Questions for authors
  • Given that review/approval accounts for 82.8% of human-assigned steps in the GDPval analysis, to what extent do these steps require human judgment? Could the authors include a multi-agent baseline in which a reviewer agent handles verification, while keeping human involvement for user-specific judgments? Comparing task performance and human effort under matched computational budgets would help clarify whether this level of human involvement is necessary.
4 comments

We thank the reviewer for the careful reading and appreciate that the reviewer sees the paper as addressing an important problem in human–agent collaboration and finding experimental evaluation reasonably comprehensive. We address each concern below.

The human worker in the main execution experiments is simulated by GeminiCLIAgent. The share of human-assigned steps increases from 15.6% to 32.9% in the user study without a corresponding comparison of "human" time or workload.

Necessity of simulated experiments: For simulated experiments, our goal is to compare orchestration policies head-to-head. That needs a controlled testbed where the task, the hidden user information, and the human worker are identical across conditions, and only the division of work changes. Real people cannot provide this: the same person cannot execute the same task under several plans without learning or fatigue effects.

We also constrain the simulated human so that it behaves like a bounded collaborator rather than a free extra agent (Appendix F):

  • It acts only on steps explicitly assigned to it
  • Each user step has a 3-minute timeout contrasting to 15 minutes for agent steps, to model bounded human effort;
  • On underspecified tasks, the missing requirements are available only to the simulated user, so the orchestrator gets them only by involving the user well.

On human workload: Our user study gives direct evidence that real users judged the extra involvement appropriate (Section 6):

  • Involvement Fit, which asks whether "the plan assigns me tasks that require human judgment," was one of the two highest-rated items (3.96/5; 85.5% agree or strongly agree). Autonomy ("I feel like I have meaningful control over how the plan is carried out") was 3.73/5, and both are significantly above neutral (p < .001, Holm-corrected).
  • Behaviorally, participants completed 86.2% of their assigned steps; unfinished steps are predominantly final review or confirmation steps where participants choose to leave with the existing output.

The plan-quality rubric was iteratively refined using only 20 human-annotated comparisons, with no reported held-out human validation.

We agree that a held-out human validation set during rubric development would have been better. Nonetheless, several pieces of evidence from the paper verifies the effectiveness of the rubric:

  1. Judge transfer. We train with Qwen3-235B-A22B-Instruct as the judge but evaluate with gemini-3-flash, a different model family (Table 1). On 200 plans sampled from different policies, scores from three judges from three model families (Qwen3-235B-A22B-Instruct, gemini-3-flash, claude-sonnet-4.6) correlate significantly (Pearson r = 0.50–0.64, all p < 1e-12; Appendix B.4).
  2. Transfer to real users' perception. The user study's four plan-quality items mirror the four rubric dimensions. Real users rated CollabOrchestra plans well above neutral on all of them: Involvement Fit 3.96, Workflow Coherence 3.96, Handoff/Deliverable Quality 3.84, Granularity Fit 3.76 (all p < .001). This is a held-out human signal on new tasks that users brought themselves, from people who never saw the rubric.

Given that review/approval accounts for 82.8% of human-assigned steps in the GDPval analysis, to what extent do these steps require human judgment? Could the authors include a multi-agent baseline in which a reviewer agent handles verification, while keeping human involvement for user-specific judgments? Comparing task performance and human effort under matched computational budgets would help clarify whether this level of human involvement is necessary.

Thank you for this suggestion. We implemented the proposed baseline and conducted a controlled experiment as follows. The results show that when users possess information unavailable to the agents, review and approval steps require meaningful human judgment; replacing human review with agent review eliminates most of the benefits of human involvement.

Experiment protocol. We take the same workflow plans produced by our orchestrator and compare three ways of executing the steps assigned to the "user":

  • Human Execution: all "user" steps are executed by the user simulator, as in Table 2.
  • Agent Review: "user" review/approval steps go to an additional GeminiCLI agent. All other "user" steps stay with the user simulator.
  • Full Agent Substitution: all "user" steps go to the additional GeminiCLI agent that receives the original task prompt, the full workflow plan, the current step's goal and deliverables, and the outcomes of previous steps. It does not receive the user profile or the user-only information for underspecified tasks. Each of its steps has a 15-minute timeout, compared with 3 minutes for the user simulator.

Results. Effect of substituting human judgment with agent execution. All conditions use identical workflow plans generated by \name and differ only in how human-assigned steps are executed. * indicates Task Performance significantly outperforms other conditions in the same Orchestrator group.

OrchestratorUser Step ExecutorGDPval Delivery RateGDPval Task Perf.GDPval-underspecified Delivery RateGDPval-underspecified Task Perf.
Qwen3-4B CollabOrchestraHuman Execution0.9700.7410.9520.645*
Qwen3-4B CollabOrchestraAgent Review0.9590.7340.9800.520
Qwen3-4B CollabOrchestraFull Agent Substitution0.9640.7420.9800.472
Qwen3-30B-A3B CollabOrchestraHuman Execution0.9570.5970.9520.474
Qwen3-30B-A3B CollabOrchestraAgent Review0.9900.7600.9850.491
Qwen3-30B-A3B CollabOrchestraFull Agent Substitution1.0000.7860.9850.481
gpt-oss-120b CollabOrchestraHuman Execution0.9260.7440.9300.703*
gpt-oss-120b CollabOrchestraAgent Review0.9950.7680.9800.536
gpt-oss-120b CollabOrchestraFull Agent Substitution0.9740.7530.9850.495

The controlled experiment reveals two key results:

  1. When agents have access to the complete task description, review steps primarily serve as verification. In this setting, an agent reviewer with a larger budget performs as well as or better than the simulated user.
  2. When users possess information unavailable to the agents, review and approval steps require meaningful human judgment. Replacing human review with agent review eliminates most of the benefits of human involvement. Specifically, on Qwen3-4B CollabOrchestra, Task Performance drops from 0.645 to 0.520, erasing 72% of the performance gain over the Agent-Only baseline (0.472). Similarly, on gpt-oss-120b CollabOrchestra, performance drops from 0.703 to 0.536, erasing 80% of the gain over Agent Only (0.495).

Since real-world tasks often involve user-specific knowledge and preferences not fully captured in the initial task description, our results underscore the importance of meaningful human involvement beyond procedural review and approval.

Beyond task performance, we want to stress that reviewing steps plays an important role to keep the human informed of what the agents did and able to step in.

  • Real users cannot read all upstream outputs, and review checkpoints are how they stay aware of what agents have done and keep the ability to intervene (the "irony of automation" we discuss in Introduction, and also Laban et al., 2026; Padmakumar et al., 2026; Guo et al., 2024 we cite in Related Work).
  • Our user study also supports this. Participants reported understanding what the agents were doing (3.75/5) and having meaningful control (3.73/5), and one participant wrote that compared with "the black box feel of traditional AI, this AI provides clearer steps, which makes it more trustworthy" (P23).

Hi Reviewer m76n,

Thank you again for your thoughtful feedback! We have posted our rebuttal addressing the concerns raised in the review and revised the manuscript (changes marked in blue) to include Section 5.3 “Is Human Involvement Necessary in Human-Agent Teaming” based on your helpful suggestions.

We would greatly appreciate it if you could take a look at our rebuttal and reconsider your assessment. Please let us know if any concerns remain or if further clarification would be helpful.

Best regards,

Submission 13806 Authors

Comments on the paper

1 comment

We thank the reviewers for their thoughtful and constructive feedback, and for recognizing the following strengths and contributions of our work:

  1. Importance of the problem and workflow-level collaboration framework. “The work addresses an important challenge in human-agent collaboration” and introduces”a practical abstraction for long-horizon occupational tasks”. (Reviewer m76n, Reviewer oCYK, Reviewer UYra)
  2. Comprehensive empirical evaluation. The evaluation covers multiple settings and provides strong empirical evidence supporting the proposed approach. (Reviewer m76n, Reviewer oCYK)
  3. Practical and well-motivated training recipe. “The combination of execution-grounded capability profiling and multi-turn reinforcement learning provides a practical approach to training orchestrators while avoiding the cost of end-to-end workflow execution in every training rollout.” (Reviewer oCYK, Reviewer UYra)

In response to the reviewer’s concern, we revise the manuscript in three key areas:

  1. We added confidence intervals and statistical significance tests to our experimental results.
  2. We clarify the design of the evaluation setup. The simulated experiments enable controlled, scalable comparisons of orchestration strategies and task outcomes, while the user study evaluates how real users perceive and participate in AI-orchestrated collaboration at the workflow level.
  3. We add Section 5.3 “Is Human Involvement Necessary in Human-Agent Teaming” to include the comparison with multi-agent baselines. Our results show that agents can effectively fulfill human-assigned responsibilities when sufficient task context is available; however, when users possess information unavailable to agents, human participation can substantially improve task performance. Since real-world tasks often involve user-specific knowledge and preferences not fully captured in the initial task description, these findings highlight the importance of meaningful human involvement and the value of the CollabOrchestra method. We thank the reviewer for suggesting this additional comparison.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.