Human-AI Agents: Collaborative Systems

• Human-AI Agents are sociotechnical systems where humans and AI cooperate in perception, reasoning, and action, ensuring human accountability.
• They encompass varied collaboration regimes—including embodied assistance, delegated and adaptive control, and bidirectional partnerships—to support shared cognitive tasks.
• Their architectures leverage API-calling, multimodal perception-language coupling, and identity-grounded frameworks to enable controlled, context-aware autonomy.

Human-AI Agents (HAIs) are sociotechnical systems in which humans and AI share perception, reasoning, action, supervision, or governance rather than reducing interaction to either passive tool use or full automation. Across recent work, this umbrella includes embodied assistive agents for physical procedures, delegated software agents that act through APIs under human authority, mixed-initiative control systems, human-human-AI triads, workflow-level economic coalitions, and long-horizon generative agents whose identity is explicitly grounded; in one paper, “HAis” is also used more narrowly for ID-RAG-enabled generative agents aligned to a real-world individual or organizational entity (Bellos et al., 24 Jul 2025, Desai et al., 25 Feb 2025, Zhu, 9 Jun 2026, Platnick et al., 29 Sep 2025).

1. Conceptual scope

A recurring definition in the literature treats AI agents as software that “perceives the world, has reasoning capabilities, and can act autonomously to carry out a user’s instructions,” with API-calling identified as the threshold that makes an LLM system genuinely “agentic” because it can trigger side effects outside itself (Desai et al., 25 Feb 2025). This immediately distinguishes HAIs from earlier generative systems that only answered, described, or recommended.

Human-centered formulations, however, reject the idea that AI agents become independent principals. The strongest normative line is that humans remain responsible for AI-agent actions, while AI systems operate with bounded initiative under human goals, platform rules, authentication requirements, and institutional constraints; the human-AI relation is therefore closer to accountable delegation or supervised tool use than to legal agency or personhood (Desai et al., 25 Feb 2025). In parallel, human-centered HCAI/HAII frameworks model the human and AI as a joint cognitive system with shared situation awareness, collaborative decision-making, collaborative control, and human final authority. The associated theory stack includes the Human-AI Joint Cognitive System (HJCS), Agent Teaming Situation Awareness (ATSA), and Shared Social Understanding (SSU), all of which treat the relevant unit of analysis as the human-AI team rather than the isolated model (Gao et al., 16 Jan 2026).

A more formal embodiment-oriented line defines coexistence in a system $S=\{A,H,E\}$ with observer-dependent quality $Q_O(t)$. On this view, a co-existing embodied agent is one that sustains meaningful and reciprocal interactions with humans and environment over time, so that long-run system quality is at least as good as the no-agent baseline (Kuehn et al., 7 Feb 2025). This formulation is notable because it shifts the target from one-shot task success to longitudinal mutual adaptation.

Another conceptual extension appears in work on identity-grounded generative agents. There, HAis are defined as generative agents whose self-perception is grounded in an explicit identity model—a Chronicle knowledge graph of beliefs, traits, values, preferences, and goals—rather than left implicit in a monolithic long-term memory stream (Platnick et al., 29 Sep 2025). This narrower usage coexists with the broader umbrella meaning of HAIs in the rest of the literature.

2. Major forms of human-AI agency

The literature now spans several distinct collaboration regimes. In embodied assistive HAIs, the human remains the embodied executor while the AI serves as a perceptual and conversational guide. The clearest instance is an AR-equipped assistant for physical procedures such as making tea, preparing pinwheel sandwiches, preparing a dessert quesadilla, and applying a tourniquet; the AI monitors progress, provides stepwise instruction, detects deviations, and enters conversation mode when timing or perception signals indicate a problem (Bellos et al., 24 Jul 2025).

Delegated software agents form a second regime. Here, the AI can book travel, post content, deploy code, purchase goods, or navigate websites, but it does so through API-mediated, authenticated, rate-limited infrastructures. The human sets goals, the AI executes with bounded discretion, and humans, firms, and platforms retain accountability (Desai et al., 25 Feb 2025).

A third regime is adaptive delegation. In a Gridworld proof of concept, both the human-like agent and the AI agent are error-prone, and an intermediary manager learns at each state which one should act next. This reframes mixed initiative as a learned control-allocation problem rather than a fixed “human leads” or “AI automates” rule (Fuchs et al., 2022).

A fourth regime is human-human-AI triadic collaboration. In programming education, the AI is not a substitute peer but an additional collaborator embedded in a human pair. The distinction between a shared AI visible to both humans and private personal AIs turns out to be consequential because shared visibility makes AI use socially observable and accountable (Daryanto et al., 17 Jan 2026).

A fifth regime is bidirectional partnership, in which support flows in both directions. In brain-tumour MRI assessment, radiologists supported by AI improved, but AI supported by radiologists also improved, and the best overall balanced accuracy came from the AI agent supported by a human one (Ruffle et al., 13 Dec 2025).

A sixth regime is longitudinal social partnership. In a six-month physical-activity study, an LLM-driven Simulated Exercising Peer (SEP) was designed not as an authoritative coach but as a horizontal workout buddy. Human peers generated stronger social presence, whereas AI peers generated steadier encouragement and stronger working alliance, producing what the authors call a “partnership paradox” (Silacci et al., 2 Feb 2026).

A final relational variant reverses the usual direction of help: humans helping AI. In a controlled experiment, helping an AI agent reduced loneliness, and this effect strengthened when the interaction supported competence and autonomy, suggesting that reciprocity in HAIs can affect the human helper’s well-being rather than only the AI system’s performance (Zhu et al., 5 Feb 2025).

Regime	Human role	AI role
Embodied assistance	Physical executor	AR guide, monitor, conversational helper
Delegated software action	Goal setter, supervisor	API-calling task executor
Adaptive delegation	Co-acting controller	Co-acting controller selected by manager
Triadic collaboration	Peer learner/collaborator	Additional visible or personal collaborator
Bidirectional partnership	Expert supporter and supported agent	Model supporter and supported agent
Social partnership	Exercising peer	Simulated exercising peer

3. Architectures and technical mechanisms

Embodied assistive HAIs are already implemented as modular, real-time systems. One architecture centers on a Conductor Process that builds a task graph from a Task Library and orchestrates multimodal components including a Data Manager, Active Perception Process, LLM Process, ASR, TTS, Timer Service, and Logger Process, with inter-process communication via shared memory and ZeroMQ. In AI-guided sessions the user wears a Microsoft HoloLens 2, with egocentric RGB, side cameras, depth, IMU, and audio, while a GoPro Hero 12 Black records exocentric video; the system was deployed on a Lenovo ThinkPad P16 Gen 2 with an NVIDIA RTX 5000 GPU (Bellos et al., 24 Jul 2025).

Perception-language coupling is a common architectural theme. In the same assistive setting, guidance generation is not end-to-end but a coordinated loop: active perception is conditioned on the current task node, out-of-sequence actions raise alerts, and the agent can shift into conversational assistance when the user stalls or deviates. Two perception options are described: a zero-shot scene-description pipeline using DINO, BLIP-2 or LaViLa, and GPT-3.5-turbo, and a supervised ResNet-based step classifier (Bellos et al., 24 Jul 2025).

Identity-grounded HAIs introduce a different architectural axis. Their core equation is $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$, where working memory is augmented with identity knowledge retrieved from a Chronicle graph $\mathcal{C}_t=(V_t,E_t)$. The decision loop explicitly separates perception, episodic retrieval, working-memory construction, identity-query formulation, identity retrieval, context augmentation, action generation, and optional identity update. In the reported implementation, the Chronicle is a NetworkX graph of relational triplets, queried by an LLM-generated relation-priority search plan and textualized back into the prompt (Platnick et al., 29 Sep 2025).

Policy-aware organizational HAIs add another layer: stakeholder mediation. HADA wraps tools and models in role-specific stakeholder agents—business, data-science, audit, ethics, and customer—coordinated by a controller agent and connected to a tools layer through A2A, MCP, and REST/GRPC façades. Alignment artifacts are stored in catalogues such as the Business-Target Catalogue, Model Catalogue, Values Catalogue, AI Metadata, Ethics Tracker, and immutable decision ledger; the concrete proof of concept exposes a credit-scoring service at POST /getLoanDecision/{modelId} (Pitkäranta et al., 1 Jun 2025).

Adaptive allocation frameworks formalize the autonomy spectrum itself. HAAS represents each subtask by five auditable cognitive dimensions—repetitiveness, technical depth, creativity, ambiguity, and human interaction—and selects among five collaboration modes: Human-Only, Copilot, Peer, Supervised, and Autonomous. A rule-based PolicyEngine enforces governance constraints before any learning occurs, and a contextual bandit then chooses among feasible modes only (Pelechanoa et al., 4 May 2026).

Implicit personalization extends HAI architectures into physiological sensing. A hybrid BCI detects subjective interest from EEG, pupil dilation, and gaze time; a graph-based TAG module extrapolates these preferences to unseen objects; and a Double DQN driving agent receives both explicit safety reward and interest-derived reward, learning to maintain safe distance while slowing near objects of interest (Shih et al., 2017).

4. Evaluation paradigms and empirical results

A foundational methodological result is that better isolated AI does not necessarily mean better human-AI teaming. In the GuessWhich cooperative game, improvements that appeared when an answering bot was paired with another AI questioner did not translate into improved human-AI team performance. Human-Alice teams using the supervised model achieved MR $= 6.86 \pm 0.53$ and MRR $= 0.27 \pm 0.03$, while the RL-fine-tuned version achieved MR $= 7.19 \pm 0.55$ and MRR $= 0.25 \pm 0.03$, with a Mann-Whitney $p = 0.44$; the broader lesson is that human-AI evaluation must be interactive, downstream-task-based, and team-centered (Chattopadhyay et al., 2017).

Embodied assistive evaluation makes this principle concrete. For first attempts with no prior training, AI guidance in physical tasks produced a Macro Success Rate of 70.00%, versus 20.00% for Unassisted and 28.57% for Paper Instructions; Step Error Rate was 16.43% for AI, versus 38.75% for Unassisted and 18.37% for Paper Instructions; completion time was slower at 186.54 seconds, versus 106.26 and 144.29 seconds. The same study introduced M-SR, $\mu$-SR, Step Error Rate, error reduction, step-guidance alignment, user-rated clarity, proactivity, ease of use, satisfaction, relevance, overall score, and cost-controlled evaluation. It also reported a multimodal dataset of 144 sessions, $Q_O(t)$0 minutes average duration, and 15.15 hours of valid exocentric recording (Bellos et al., 24 Jul 2025).

Educational triads show a different pattern. In human-human-AI programming, the proportion of AI-generated code was 23.1% in the HAI baseline, but only 1.4% in HHAI-Shared and 1.2% in HHAI-Personal, with Wald $Q_O(t)$1. Responsibility for understanding AI suggestions was highest in HHAI-Shared at $Q_O(t)$2, compared with 2.80 in HAI and 2.60 in HHAI-Personal, and collaborative learning and social presence both increased significantly relative to HAI. This supports the claim that shared visibility of AI use activates socially shared regulation rather than simple reliance (Daryanto et al., 17 Jan 2026).

Bidirectional clinical partnerships show that support can improve both agents. In MRI-guided brain-tumour assessment, radiologist-alone balanced accuracy was 0.698 and radiologist-with-model support was 0.743, while model-alone was 0.824 and model-with-radiologist support was 0.841. Confidence-accuracy correlation rose from $Q_O(t)$3 without support to $Q_O(t)$4 with support, and Cohen’s kappa increased from 0.338 to 0.484, indicating gains not only in accuracy but also in metacognitive calibration and consistency (Ruffle et al., 13 Dec 2025).

Operations benchmarks extend evaluation to workflow economics. In InventoryBench, over 1,320 instances, Gemini 3 Flash achieved normalized reward 0.445 with OR alone, 0.494 with LLM alone, 0.538 with OR$Q_O(t)$5LLM, and 0.501 with LLM$Q_O(t)$6OR. In the associated human study, Mode B (OR$Q_O(t)$7LLM$Q_O(t)$8Human) achieved 0.534 $Q_O(t)$9, outperforming both Mode A (OR$A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$0Human) at 0.466 $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$1 and OR$A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$2LLM alone at 0.482 $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$3. The paper also derived a distribution-free lower bound implying that at least 20.3% of individuals had strictly positive individual-level complementarity, with 95% bootstrap confidence interval 0.147, 0.329.

Longitudinal social partnership produces still another evaluation profile. In the six-month workout-buddy trial, social presence differed by condition, $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$4, with HUM exceeding both SEPH and SEPC, whereas working-alliance bond differed more strongly, $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$5, with both AI-peer conditions exceeding HUM. The result was a stable separation between authentic social presence and reliable alliance rather than a general AI-over-human advantage (Silacci et al., 2 Feb 2026).

5. Governance, accountability, and economic attribution

Governance-oriented HAI research insists that agency does not dissolve human responsibility. Responsible AI Agents argues that humans are responsible for AI-agent actions, that software should not be granted legal personhood, and that APIs, permissions, confirmations, authentication, reversibility, logging, and ex ante explanations are the operative mechanisms for constraining delegated action. It also extends value alignment beyond “helpful, honest, harmless” toward loyalty, conflict-of-interest avoidance, clarification under ambiguity, and user contestation (Desai et al., 25 Feb 2025).

Organizationally, HADA turns this into a multi-agent governance fabric. Role-specific stakeholder agents expose conversational interfaces so business, data-science, audit, ethics, and customer actors can query, steer, audit, or contest decisions across strategic, tactical, and real-time horizons. In the retail-bank prototype, yearly OKRs flow into quarterly KPI changes, model versions are approved or blocked through role-specific privileges, and customer challenge to ZIP-code use triggers ethics ticket ETH-512, sensitive-attribute flagging, and retraining (Pitkäranta et al., 1 Jun 2025).

Adaptive autonomy work treats governance as a tunable variable rather than a binary switch. HAAS shows that tightening policy does not primarily replace AI with human labor; instead it predictably converts Autonomous assignments into Supervised collaboration. In manufacturing, stronger governance improved both objective and fatigue up to L3, with cost per batch moving from 147.13 at L0 to 139.07 at L3 and fatigue from 0.926 to 0.842, a workload-buffering effect that contradicts the view of governance as pure overhead (Pelechanoa et al., 4 May 2026).

Economic theories of HAIs then move from governance to valuation. Agentomics models the workflow as an assignment $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$6 of human and AI agents to stages, with net workflow value $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$7, coalition value $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$8, and Shapley attribution $A_t=\Pi(WM_t \oplus K_t^{\text{ID}})$9 for AI agents. In the security-operations case study, the grand coalition raised net value to $\mathcal{C}_t=(V_t,E_t)$02,845, and <a href="https://www.emergentmind.com/topics/shapley-values" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Shapley values</a> were $\mathcal{C}_t=(V_t,E_t)$1620 for the <a href="https://www.emergentmind.com/topics/triage" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Triage</a> Agent, and $285 for the Investigation Agent, yielding a normative pricing benchmark $\mathcal{C}_t=(V_t,E_t)$2 (Zhu, 9 Jun 2026).

6. Limitations, controversies, and research directions

Several limitations recur across the literature. Embodied assistive HAIs are often evaluated on small studies: one influential procedural benchmark used only 12 participants, 4 tasks, and data collected in a single room, with dialogue relevance the weakest user-rated dimension and intervention mechanisms that were rule-triggered rather than uncertainty-aware (Bellos et al., 24 Jul 2025). Adaptive delegation work still relies on simulated humans in Gridworld, injected error states, coarse trajectory-level reward, and zero switching cost, so it remains a foundational demonstration rather than a deployment-ready theory of mixed-initiative control (Fuchs et al., 2022). Identity-grounded HAIs currently use small, handcrafted Chronicles, partial ID-RAG implementations, and no deployed action-validation or dynamic identity-update machinery (Platnick et al., 29 Sep 2025).

A second controversy concerns what counts as “better” AI. The cooperative-game result that better AI-AI performance does not predict better human-AI performance remains a caution against proxy benchmarks (Chattopadhyay et al., 2017). The workout-peer trial sharpens this further: AI peers were valuable not by achieving human-like authenticity, but by offering consistent, low-stakes support; attempts to imitate human authenticity too closely risk disengagement (Silacci et al., 2 Feb 2026). This suggests that anthropomorphism is not a universal design objective.

A third open issue is the status of simulation. Platforms such as Humanoid Agents enrich generative-agent simulation with basic needs, emotion, and relationship closeness, showing systematic adaptation in routines and dialogue, but they still fall short of direct human-in-the-loop HAI evaluation and are better characterized as simulation infrastructures than as full HAI testbeds (Wang et al., 2023). By contrast, coexistence-oriented work argues that long-term HAIs should be situated, mutable, and open-ended, sustaining meaningful and reciprocal interactions in a system $\mathcal{C}_t=(V_t,E_t)$3 rather than merely optimizing fixed tasks (Kuehn et al., 7 Feb 2025).

Future directions are correspondingly broad. Human-centered HCAI/HAII surveys call for deeper theories of dynamic control transfer, trust evolution, cross-cultural adaptation, and controllability for generative and embodied AI, together with panoramic evaluation across technology, organization, and society (Gao et al., 16 Jan 2026). Prosociality studies suggest that reciprocal designs in which humans help AI can reduce loneliness when competence and autonomy are supported, but also warn against fake neediness and manipulative dependency (Zhu et al., 5 Feb 2025). Taken together, the literature points toward HAIs that are multimodal, policy-aware, identity-grounded, economically auditable, and longitudinally evaluated—yet also explicit about bounded authority, human accountability, and the irreducibility of local social context.