Human-Agent Teaming

• Human-Agent Teaming is a collaborative paradigm where humans and autonomous agents operate as equal partners, combining contextual insight with computational precision.
• These systems employ functionalities such as shared situation awareness, explainability, and proactive communication to maintain robust team dynamics across complex environments.
• Practical applications, including UAV surveillance and ground-robot search, demonstrate enhanced performance through adaptive function allocation and formalized work agreements.

Human–Agent Teaming (HAT) refers to configurations in which humans and autonomous artificial agents operate as equal partners to jointly pursue shared objectives. This paradigm departs from traditional human–machine interaction—where AI systems are mere tools receiving continuous direction from humans—by establishing autonomous agents as active team members with social, cognitive, and communicative competencies. Core to HAT is the vision of human and agent complementarity: humans supply contextual knowledge, experience, and decision-making under uncertainty, while agents provide speed, precision, and persistent attention across distributed environments. Modern frameworks emphasize modularity, explainability, proactive communication, and adaptive function allocation to support fluid team dynamics in high-stakes, real-world domains.

1. Foundations and Dimensions of Human–Agent Teaming

The foundational analysis of human–agent teams is articulated along five core dimensions (Diggelen et al., 2019):

1. Environment: The nature of the operational context (dynamic/static, predictable/unpredictable) critically shapes the need for shared situation awareness. Increased environmental uncertainty complicates information pooling across the team.
2. Mission/Task Properties: Duration, risk, interdependency, and time criticality define what levels of control and information flow are necessary. High time pressure/risk mandates rapid transitions between "out-of-the-loop" and "in-the-loop" control regimes.
3. Team Organization: Characteristics such as team size, distribution, skill/authority differentiation, and physical proximity determine coordination requirements and trust calibration mechanisms.
4. Team Dynamics: Temporal structure (standing, ad hoc, evolving teams) and lifecycle phase (commissioning, action, debriefing) affect the formation and updating of shared mental models for coordinated action.
5. Communication: The modalities and infrastructure for information exchange (richness, mapping, reliability) are central to maintaining and restoring common ground, especially as complexity or adversity increases.

Illustrative scenarios (e.g., UAV surveillance, ground-robot search) anchor these dimensions, underscoring the multidimensional variability of teaming requirements.

2. Core Teaming Functionalities

A general framework for HAT is defined by a small set of functional building blocks, each designed to be modular and pluggable to suit various autonomous (task-oriented) systems (Diggelen et al., 2019):

• Shared Situation Awareness: Mechanisms for synchronizing each agent's perception of the environment, status, and team progress. This includes explicit "intention awareness," ensuring that agents selectively share high-value information.
• Explainable and Explanatory AI: Agents not only execute actions but provide justifications (the "why"), supporting trust, predictability, and transparency.
• Interpredictability: Agents predict (and are predictable to) the actions and needs of teammates. Team training (procedural/cross-training for shared mental models) is vital for effective interpredictability.
• Proactive Communication: Agents initiate information transfer when contextually indicated rather than solely upon human query, considering teammate cognitive state and preferences.
• Directability: Agents employ and respect explicit work agreements—contractual constraints that delineate obligations and prohibitions. These agreements facilitate dynamic task allocation and can be specified via formal deontic logic. An example formalization:

$$O(\text{notify}(\text{BaseCommander}, \text{HostileContact})) \iff \text{detect}(\text{HostileContact}) = \text{true}$$

Collectively, these functionalities underpin effective collaboration without demanding systemic redesign.

3. SAIL: Technical Architecture and Implementation

The Social Artificial Intelligence Layer (SAIL) (Diggelen et al., 2019) embodies these teaming functionalities in a modular middleware architecture:

• Component Structure: SAIL consists of (i) humans interfaced through flexible UIs, (ii) task-oriented AI modules for low-level execution, and (iii) Social AI modules that abstract teaming functionalities and mediate between humans and the task-level AI.
• Communication Protocol: The Human–Agent Teaming Communication Language (HATCL) is a FIPA-ACL-inspired protocol, structuring message exchange as:

$$\langle \text{Performative}, \text{Sender}, \text{Receiver}, \text{Content}, \ldots \rangle$$

with performatives capturing intention (Inform, Query, Propose, etc.), and semantic anchors grounding high-level commands into executable system variables (mapping, for instance, a "no left-turn" agreement into navigation constraints).

• Ontological Layer: A dual-layer ontology supports unambiguous interpretation: a domain-independent layer (Actor, Plan, Goal, Action) and a domain-specific layer mapped via semantic anchors.
• Implementation Platform: SAIL is implemented atop Akka for distributed, cross-platform modularity, facilitating flexible deployment across heterogeneous agents and operating environments.
• Mechanism of Semantic Anchoring: Team agreements expressed in HATCL are algorithmically mapped to concrete code-level changes; e.g.,
  • Work agreement "prohibited(turn_left)" $\rightarrow$ code: tl_permitted = false or adjustment to a navigation desirability map.

4. Prototype Application and Evaluation

A proof-of-concept prototype utilizing SAIL demonstrates HAT in a military surveillance scenario with UAVs in Gazebo simulation (Diggelen et al., 2019):

• Scenario Description: Multiple unmanned aerial vehicles (UAVs) with autonomous navigation and object detection support a human base commander.
• SAIL Integration: The TAI modules of the UAVs are extended with SAI to realize:
  • Shared awareness: critical updates pushed only upon anomalies
  • Proactive communication: "ProCom" module modulates information flow by estimating event significance and operator cognitive load
  • Enforcement/monitoring of work agreements through HATCL
• Human-Interface: A multimodal avatar acts as an intermediary, delivering compressed situational updates (visual, speech, text, touch), following a management-by-exception principle to minimize distraction.
• Performance Observation: While direct metrics are not reported, the architecture enables dynamic team coordination, trust calibration, and rapid transition between control regimes as required by environmental conditions and mission parameters.

5. Formalization and Cross-cutting Integration

Formal representations play a crucial role in encoding and transferring team agreements:

• Deontic Logic for Work Agreements: Obligations and prohibitions are encoded formally, ensuring enforceable commitments at the agent level.
• Semantic Anchoring: Abstract communicative acts in HATCL are programmatically mapped into agent-executable behaviors, democratizing the injection of high-level team intent into heterogeneously designed task-level AI modules.
• Ontological Structuring: The synergy of generic and domain ontologies permits scalable and interpretable interpretation of messages and agreements, crucial for robust and safe cross-domain deployments.

6. Future Research Directions

Three primary trajectories are identified for advancing HAT (Diggelen et al., 2019):

• Enhanced Semantic Anchoring: More sophisticated mechanisms are needed to bridge abstract agreements with diverse underlying architectures, especially neural systems relying on distributed representations.
• Validation Over Extended Team Lifecycles: Longitudinal studies are needed to assess HAT efficacy in terms of team trust, learning, adaptability, and evolving group dynamics.
• Advanced Interaction Modalities: The integration of immersive techniques (e.g., tele-presence) is proposed to further blur boundaries between human and agent team members, pushing toward more naturalistic and adaptive teaming.

7. Synthesis and Outlook

Human–Agent Teaming, as operationalized in the SAIL framework, signifies a fundamental reimagining of human–machine interaction. By codifying modular teaming functionalities (shared awareness, explainability, interpredictability, proactive communication, directability) as pluggable services, HAT systems accrue flexibility, scalability, and robustness necessary for complex tasks in uncertain environments. The explicit formalization of work agreements, adoption of ontological mappings, and real-world prototype validation advance the domain toward deployable, equal-partner human–agent teams. Critical open research problems remain: strengthening the formal grounding of team agreements in dynamic, distributed agent architectures and developing validation protocols that capture the long-term, adaptive, and sociotechnical facets of HAT. Collectively, these avenues define the frontier for next-generation collaborative autonomous systems.