HILA: Human-In-the-Loop Multi-Agent Collaboration

• HILA is a paradigm where multiple AI agents and humans collaborate directly through intervention, supervision, and co-planning to enhance decision-making and safety.
• Key architectural patterns include decoupled oversight, protocol substrates, and shared cognitive spaces that enable transparent and coordinated interactions.
• Empirical evaluations in software, healthcare, robotics, and other fields show improved performance and adaptability while highlighting challenges in latency and expert dependency.

Searching arXiv for papers on human-in-the-loop multi-agent collaboration and closely related protocols/frameworks. Human-In-the-Loop Multi-Agent Collaboration (HILA) denotes a class of systems in which multiple AI agents and one or more humans collaborate through explicit intervention, supervision, correction, arbitration, or co-planning, rather than treating the human as a final approver outside the workflow. In recent research, HILA appears as a response to three recurrent limitations: purely autonomous multi-agent systems remain “closed-world” and brittle on tasks beyond their knowledge horizon, outcome-level supervision exposes only final outputs while hiding coordination failures in intermediate reasoning, and embedded approval logic does not scale across heterogeneous agent environments (Yang et al., 9 Mar 2026, He et al., 21 May 2026, Cheng et al., 24 Apr 2026). The term therefore covers protocol substrates, oversight architectures, learning frameworks, and application systems in which humans participate as supervisors, validators, escalation authorities, collaborators, teachers, or direct task participants.

1. Conceptual scope and distinguishing features

HILA is defined less by a single algorithm than by a recurring control pattern: multiple agents collaborate on a task, but the human remains inside the operative loop at decision points that matter for safety, correctness, alignment, or capability growth. In the literature, this pattern is explicitly contrasted with “pointwise and reactive” interaction, where users approve or correct individual actions without visibility into downstream consequences, and with traditional Human-in-the-Loop designs that treat the human as an external validator rather than a genuine collaborator in a multi-agent process (He et al., 12 Mar 2026, Melih et al., 28 Oct 2025). It is also contrasted with AI-centered systems in which human oversight is embedded ad hoc inside each workflow rather than treated as a reusable system component (Cheng et al., 24 Apr 2026).

A second distinguishing feature is that HILA is not confined to deployment-time approval. Some systems place the human in the loop during training, as in single-human intervention for collaborative robot policy learning or in policy-corrected demonstrations for reinforcement learning; others place the human in iterative execution-time replanning, as in ventilator decision support; still others use expert review for data curation, as in multi-agent mathematical dataset construction (Ji et al., 2024, Islam et al., 2023, Li et al., 22 May 2026, Liu et al., 2 Jun 2025). This suggests that HILA is best understood as a family of collaboration regimes spanning training, planning, execution, validation, and continual improvement.

A common misconception is that HILA is equivalent to final approval gates. The recent planning and interaction literature rejects that reduction explicitly: process-level supervision is presented as necessary because failures often arise from task decomposition, dependency structure, hidden constraints, or intermediate coordination states rather than from the final output alone (He et al., 21 May 2026). A plausible implication is that HILA systems are most distinctive when they expose internal coordination artifacts—plans, states, conflicts, explanations, or candidate futures—as first-class objects for human inspection.

2. Architectural and protocol patterns

The architectural literature converges on several recurring patterns. One pattern is the decoupled oversight architecture, in which human interaction management is externalized from application logic into a shared control-plane component. In that design, agents issue structured requests such as /api/hitl/request, the HITL component evaluates intervention conditions, resolves the appropriate human participant, routes the interaction through channels such as Slack, Email, SMS, or an AMP UI portal, and returns the decision through polling or callback. The corresponding design framework is organized along four dimensions—intervention conditions, role resolution, interaction semantics, and communication channel—formalized as WHEN, WHO, WHAT, and WHERE (Cheng et al., 24 Apr 2026).

A second pattern is the protocol substrate for interoperable coordination. “MPAC: A Multi-Principal Agent Coordination Protocol for Interoperable Multi-Agent Collaboration” is explicitly positioned as an application-layer protocol for settings where agents act on behalf of independent principals rather than a single owner. MPAC introduces five layers—Session, Intent, Operation, Conflict, and Governance—makes intent declaration a precondition for action, represents conflicts as first-class structured objects, and supports human-in-the-loop arbitration through a pluggable governance layer. The specification defines 21 message types, three state machines with normative transition tables, Lamport-clock causal watermarking, two execution models, three security profiles, and optimistic concurrency control on shared state (Qian et al., 10 Apr 2026). In HILA terms, this is a move from ad hoc coordination toward explicit protocol semantics for conflict, governance, and shared-state control.

A third pattern is the shared cognitive substrate. “Human-Machine Social Hybrid Intelligence” organizes collaboration around a Shared Cognitive Space, formalized as

$$S_t = \{ \mathcal{O}_t, \mathcal{K}, \mathcal{H}_t, \mathcal{T}_t, \mathcal{A}_t \},$$

where current objects, persistent knowledge, event history, task state, and agent state are maintained in a shared world model (Melih et al., 28 Oct 2025). Closely related ideas appear in orchestrated multi-agent planning, where the plan is a directed acyclic graph of agent-executable subtasks and dependencies, and in multimodal human–multi-robot systems, where each robot maintains local perception–planning–action loops under a team-level coordinator that regulates participation (He et al., 21 May 2026, Hasan et al., 24 Mar 2026). This suggests that shared state, rather than free-form chat alone, is a central organizing principle of modern HILA systems.

A fourth pattern is the staged human-governed workflow. In HULA for software development, the workflow is partitioned into task setup, planning, coding, and pull-request creation, with explicit human approval gates between major stages. In CollabToolBuilder, the workflow is split across Coach, Coder, Critic, and Capitalizer agents, with human pre- and post-guidance hooks at each stage (Takerngsaksiri et al., 2024, Xavier et al., 1 Dec 2025). Across both cases, human authority is exercised not only through acceptance or rejection, but also through editing intermediate artifacts and steering regeneration.

3. Forms of human participation

The role of the human varies substantially across HILA systems, but the literature is unusually explicit about these roles. In decoupled HITL architectures, humans can act as supervisor, validator, escalation authority, domain expert, compliance authority, or monitor/auditor, with role assignment determined dynamically by context (Cheng et al., 24 Apr 2026). In enterprise software development, the human is formalized as one of three agents—$\mathcal{H}$ alongside planner $\mathcal{P}$ and coder $\mathcal{C}$—and retains authority over file localization, coding-plan approval, code regeneration, and final pull-request creation (Takerngsaksiri et al., 2024). In human–MARL frameworks such as COGMENT, the human is an actor who can send actions, feedback, and recommendations through the same orchestration substrate used by other agents (Navidi et al., 2020).

Embodied and physically grounded HILA systems assign different roles again. In a multimodal human–multi-robot interaction framework, the human is primarily a conversation partner and task requester in a shared physical space; the team-level coordinator uses response-likelihood scores and thresholding to determine which robot should respond, and multiple selected agents respond sequentially rather than simultaneously (Hasan et al., 24 Mar 2026). In multi-robot collaboration from single-human guidance, one human intermittently switches control among robots during training, providing short periods of direct control that encode different team roles without requiring one demonstrator per robot (Ji et al., 2024). In multi-agent strategy explanation for human-robot collaboration, the human is less a direct controller than a collaborator who must align with the robot on a shared strategy after receiving landmark-based visual or textual explanations (Pandya et al., 2023).

The recent co-planning literature broadens the role further from supervisor to process-level collaborator. AMBIPOM formalizes human–LLM co-planning along three axes—mode, scope, and level—and supports both semantic feedback and structural edits over a visible DAG plan (He et al., 21 May 2026). Simulation-in-the-loop goes further by proposing that the human should inspect alternative future trajectories before committing to a decision, shifting collaboration from reactive supervision toward exploratory sensemaking (He et al., 12 Mar 2026). A plausible implication is that HILA increasingly treats humans as planners and arbiters of process, not merely as judges of completed outputs.

Expert review and curation constitute another human role that is central in some HILA systems. STORM-BORN uses a six-agent generation pipeline followed by mathematicians who audit quality, eliminate samples with no research value, manually modify borderline cases, and deliberately select the 100 most difficult problems from 2,000 synthetic samples (Liu et al., 2 Jun 2025). Here, the human is neither annotator in the narrow sense nor deployment-time supervisor; the human is the epistemic authority that converts machine-generated candidates into a benchmark.

4. Learning, adaptation, and control

A large part of the HILA literature concerns not only when humans intervene, but how systems learn from intervention. The most explicit formalization appears in “Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning,” which defines a metacognitive policy over three discrete actions,

$$\mathcal{A} = \{a_{\text{eval}}, a_{\text{create}}, a_{\text{defer}}\},$$

and optimizes deferral decisions through a cost-aware reward

$$R(s_t, a_t) = \begin{cases} R_{\text{gt}}(y(a_t)), & a_t = \text{EVAL} \ R_{\text{gt}}(y(a_t)) - C_{\text{create}}, & a_t = \text{CREATE} \ R_{\text{gt}}(y_{\text{human}}(a_t)) - C_{\text{defer}}, & a_t = \text{DEFER}. \end{cases}$$

The inner loop applies Group Relative Policy Optimization to learn when to defer, while the outer loop turns expert demonstrations into supervised fine-tuning signals through

$$L_{\text{SFT}}(\theta) = - \sum_{i=1}^{L} \log \pi_\theta(t_i \mid s_t, t_{1:i-1}),$$

so that human consultation becomes long-term capability growth rather than a one-off patch (Yang et al., 9 Mar 2026).

Other systems instantiate adaptation through different control primitives. VDSS maintains a clinician-specific preference state $\theta_d$ and updates it online after each accepted ventilator adjustment cycle:

$$\theta_d \leftarrow \mathrm{BanditUpdate}\!\left(\theta_d;\,x_t, \tilde a_t, h_t, p_t\right).$$

This contextual-bandit loop uses the final accepted decision, the cycle trace, and preference signals to personalize subsequent recommendations while structured rejection feedback triggers targeted replanning at the minimal decision layer—strategy selection, mode selection, or parameter planning—rather than rerunning the entire pipeline (Li et al., 22 May 2026). HARP requests human assistance only when the variance of group return exceeds the maximum value in a historical queue, then evaluates human-proposed regroupings with a Permutation Invariant Group Critic before accepting them (Hu et al., 2024).

In reinforcement-learning settings, human intervention often enters as demonstration or correction rather than explicit planning feedback. In a multi-drone airport defense simulator, policy correction—humans controlling a trained agent only when needed—produced faster learning than either human-only demonstrations or agent demonstrations, while also lowering mental demand, temporal demand, effort, and frustration relative to full human control (Islam et al., 2023). In COGMENT, the orchestrator supports combining rewards from multiple sources, generating offline datasets, and using humans and agents as actors in the same loop (Navidi et al., 2020). In single-human multi-robot guidance, collaboration is learned from 40 minutes of human guidance together with a theory-of-mind-inspired teammate model, with the PE-T variant predicting whole-team actions from one seeker’s observation (Ji et al., 2024).

Prompt- and memory-level adaptation is another important strand. CollabToolBuilder uses a Reinforced Dynamic Prompt of the form

$$\{ROLE\text{-}GOAL\text{-}CONSTRAINTS\} + \{STATE\ OBSERVATION\} + \{TASK\} + \{EXAMPLES\} + \{FEEDBACKS\},$$

and updates it with macro-level automatic scores and micro-level human or self-critique after each iteration (Xavier et al., 1 Dec 2025). This suggests that HILA adaptation often occurs above the level of gradient updates: through memory, prompt reinforcement, reusable tools, and persistent validated artifacts.

5. Representative domains and empirical evidence

The empirical literature is distributed across software engineering, healthcare, emergency response, robotics, reinforcement learning, tool building, and dataset curation. The table below organizes several representative systems.

System	Domain	Human role
HULA	Software development	Planner/code reviewer and gatekeeper
VDSS	Ventilator decision support	Clinician reviewer and preference source
HMS-HI	Urban emergency response	Human expert agent and override authority
HARP	SMAC group-oriented tasks	Deployment-time regrouping advisor
MPAC	Multi-principal coordination	Governance and arbitration participant
STORM-BORN	Math dataset curation	Expert auditor and selector

In software engineering, HULA was deployed inside Atlassian JIRA. In the online evaluation, out of 663 issues where practitioners used HULA, 527 had successful plan generation, 433 plans were approved, 376 had successful code generation, 95 were raised as PRs, and 56 were merged, yielding a Plan Approval Rate of 82%, a Raised PR Rate of 25%, a Merged PR Rate of 59% among raised PRs, and an overall 8% rate of successfully merged HULA-assisted PRs (Takerngsaksiri et al., 2024). These numbers are consistent with the paper’s broader claim that human-governed staged assistance is most effective for planning and drafting, not autonomous completion.

In healthcare, VDSS was evaluated on a multi-center ICU cohort with 1309 structured records, 7447 ventilator setting entries, 13 ventilation modes, and 2 ventilator brands. On retrospective next-step replay, GPT-5.2 used as a single model achieved MSE 0.343, average $\mathcal{H}$0 0.153, and overall expert rating 2.63, whereas GPT-5.2 within VDSS achieved MSE 0.102, average $\mathcal{H}$1 0.743, and overall rating 4.11; acceptability improved from 2.75 to 4.09 and safety from 2.43 to 4.46 (Li et al., 22 May 2026). This is one of the clearest demonstrations that modular HILA structure can outperform direct generation with the same backbone.

In high-stakes coordination, HMS-HI was validated in a high-fidelity urban emergency response simulation and reported a reduction in civilian casualties from 112.9 to 31.5 and a reduction in NASA-TLX cognitive load from 81.2 to 24.7 relative to traditional HiTL, corresponding to the paper’s headline claims of 72% fewer casualties and 70% lower cognitive load (Melih et al., 28 Oct 2025). In protocol engineering, MPAC reported a 95 percent reduction in coordination overhead and a 4.8 times wall-clock speedup versus a serialized human-mediated baseline in a controlled three-agent code review benchmark (Qian et al., 10 Apr 2026).

In multi-agent reinforcement learning and robotics, HARP achieved 100% test win rate on all six evaluated SMAC maps while keeping human participation below 25% on all maps and as low as 2.61% on 8m (Hu et al., 2024). “Enabling Multi-Robot Collaboration from Single-Human Guidance” reported improvement of the success rate of a challenging collaborative hide-and-seek task by up to 58% with only 40 minutes of human guidance, and also demonstrated transfer to real robots (Ji et al., 2024). In airport defense, policy-corrected demonstrations yielded faster learning than learning from humans or agents and reduced mental and temporal demands relative to full manual control (Islam et al., 2023).

In expert curation, STORM-BORN used a human-in-the-loop multi-agent pipeline to produce 2,000 synthetic mathematical derivation samples and deliberately selected the 100 most difficult problems; even GPT-o1 solved fewer than 5% of them, while fine-tuning on STORM-BORN improved accuracy by 7.84% for LLaMA3-8B and 9.12% for Qwen2.5-7B (Liu et al., 2 Jun 2025). This shows that HILA is not restricted to execution-time collaboration; it also functions as a method for building high-quality knowledge resources.

6. Limitations, trade-offs, and open directions

The literature is explicit that HILA does not eliminate trade-offs; it reorganizes them. One major trade-off is effort versus control versus risk. AMBIPOM’s user study reports hybrid workflows and effort-control-risk trade-offs across semantic versus structural, global versus targeted, and low- versus high-level edits (He et al., 21 May 2026). Semantic feedback is typically lower-friction but less predictable; structural editing offers greater determinism but higher manual burden. A plausible implication is that mature HILA systems will need to support both modalities simultaneously rather than selecting one.

A second trade-off is oversight quality versus latency and orchestration overhead. VDSS reports an average workflow time of 305.3 seconds per cycle on local inference, with waveform analysis alone taking 56.46 seconds and completion failure rate below 7% (Li et al., 22 May 2026). Centralized coordination can also become a bottleneck: the multimodal human–multi-robot framework notes that as the number of agents grows, arbitration complexity and latency may degrade conversational fluency (Hasan et al., 24 Mar 2026). These findings caution against assuming that more structure always yields better usability.

A third limitation concerns expert dependence and human bottlenecks. HILA’s performance can depend strongly on who the human is and what kind of feedback they provide. The metacognitive HILA framework shows that stronger experts produce better results, and most experiments still use GPT proxies for humans rather than sustained real-human collaboration (Yang et al., 9 Mar 2026). STORM-BORN depends on mathematicians for strict audits and manual optimization of borderline cases (Liu et al., 2 Jun 2025). In CollabToolBuilder, most human time is concentrated in the Coder stage, indicating that some supposedly automated loops remain human-intensive in practice (Xavier et al., 1 Dec 2025).

A fourth limitation is incomplete formalization or evaluation in parts of the literature. The decoupled HITL architecture is architectural and protocol-oriented, but provides no benchmark, user study, or controlled experiment (Cheng et al., 24 Apr 2026). HARP demonstrates strong results in SMAC but provides limited user-study detail and no dedicated ablation isolating the permutation-invariant critic from the rest of the system (Hu et al., 2024). HMS-HI reports strong simulation gains, but the authors explicitly note limits in simulation fidelity, scale, long-term adaptation, and embodiment (Melih et al., 28 Oct 2025).

A final open direction concerns the boundary between HILA as control and HILA as foresight. Simulation-in-the-loop argues that humans often have control over local actions but lack foresight into downstream consequences, and therefore need access to alternative future trajectories before commitment (He et al., 12 Mar 2026). This suggests a broader research agenda: future HILA systems may need to integrate explicit governance, shared state, metacognitive deferral, structured explanations, and simulation-based preview into a single anticipatory collaboration stack.

Taken together, the literature supports a narrow but robust conclusion: HILA is not simply human approval layered onto multi-agent systems. It is an emerging systems paradigm in which humans and agents share responsibility for decomposition, arbitration, validation, adaptation, and governance across the full lifecycle of multi-agent work.