Sensible Agent in Multi-Domain Systems

• Sensible agent is a context-aware intelligent system characterized by adaptive reasoning, persistent memory, and structurally grounded behavior across diverse domains.
• It employs cognitive loops, dynamic context management, and secure perception to ensure socially appropriate and effective interactions.
• Key frameworks like SCC, Synergy, and A2AI illustrate its practical applications in IoT, XR, game AI, and empathetic dialogue.

A sensible agent is not a single standardized construct in the recent arXiv literature, but a recurring design ideal for agents whose behavior is context-appropriate, structurally grounded, and operationally constrained rather than merely prompt-driven. In different subfields, the term has referred to cognitive BDI agents for Sense-Compute-Control applications in the Internet of Things and Services, persistent and socially legible agents on the Open Agentic Web, proactive AR agents that adapt both what assistance to offer and how to deliver it, and Go systems that optimize over score margins rather than only binary win probability (Moin, 2020, Nie et al., 30 Mar 2026, Lee et al., 11 Sep 2025, Morandin et al., 2018). Related agent frameworks strengthen the same design trajectory through ReAct-grounded execution, dynamic reasoning-context management, and environment-situated partial observability (Gao et al., 22 Aug 2025, Lian et al., 13 Apr 2026, Bentley et al., 3 Sep 2025).

1. Meanings of the term across the literature

In the IoTS literature, a sensible agent is an intelligent, cognitive agent for Sense-Compute-Control (SCC) applications whose internal cycle is sense-deliberate-act, closely matching the SCC sequence of sensing, computation, and control (Moin, 2020). In this usage, sensibility denotes autonomy, reactiveness, proactiveness, and sociality under dynamic environmental conditions.

In the Synergy architecture, the term is tied to what the paper calls an Agentic Citizen, defined by Agentic-Web-Native Collaboration, Agent Identity and Personhood, and Lifelong Evolution (Nie et al., 30 Mar 2026). Sensibility here is not primarily about one-shot task completion; it is about continuity as a social entity, persistent relationships, and experience-centered adaptation over time.

In the XR literature, Sensible Agent is the name of a framework for unobtrusive interaction with proactive AR agents. Its central claim is that the agent must adapt both “what” assistance to offer and “how” to deliver it, using real-time multimodal context sensing to remain minimally intrusive and socially appropriate (Lee et al., 11 Sep 2025).

In game-playing research, SAI expands “sensible” play beyond fixed-komi win maximization. By modeling winrate as a function of komi and allowing a configurable bias toward larger margins of victory, it aims to avoid “minimal-win” behavior and to handle handicaps more naturally (Morandin et al., 2018).

Usage	Operational emphasis	Representative mechanism
SCC/IoTS sensible agent	Cognitive autonomy in sensing and control	BDI sense-deliberate-act cycle
Synergy agent	Collaboration, identity, evolution	Typed memory, agenda, experience recall
Sensible Agent for XR	Unobtrusive proactive assistance	Multimodal context sensing and modality adaptation
SAI for Go	Score-aware, handicap-aware play	Multiple-komi sigmoid and branching self-play

This distribution suggests a family resemblance rather than a single canonical definition: sensibility is repeatedly associated with context dependence, bounded perception, persistent internal structure, and adaptation of action to task and social setting.

2. Cognitive loops, execution models, and tool-mediated action

A major lineage of sensible-agent design begins with explicit cognitive control loops. For SCC applications in the IoTS, the BDI architecture organizes the agent around Beliefs, Desires, Intentions, and Plans, with the reasoning process implemented as Sense (Perceive & Update Beliefs), Deliberate (Consider Desires, Update Intentions), and Act (Select & Execute Plans/Actions) (Moin, 2020). This formulation is presented as especially well matched to SCC workloads because it combines autonomy with environmental reactivity.

A more recent execution-centered formulation appears in AgentScope 1.0, which abstracts agentic applications into Message, Model, Memory, and Tool modules, while grounding behavior in the ReAct paradigm (Gao et al., 22 Aug 2025). AgentScope defines an agent through three core functions—reply, observe, and handle_interrupt—and formalizes interaction with environment and tools as

$$\text{For each step } t: \begin{cases} \text{Thought}_t = \text{Reason}(h_{t-1}, o_{t-1}) \ a_t, T_t = \text{Act}(\text{Thought}_t) \ o_t = E(a_t, T_t) \ h_t = h_{t-1} \cup \{(\text{Thought}_t, a_t, T_t, o_t)\} \end{cases}$$

where $h_t$ is agent history. The same framework treats agents as callable tools, supports asynchronous model and tool calls, allows dynamic activation and deactivation of tool groups, and provides a runtime sandbox plus a scalable evaluation module with a visual studio interface (Gao et al., 22 Aug 2025).

A separate but complementary execution problem is reasoning-history management. SWE-AGILE addresses the dilemma that retaining full reasoning history causes context explosion and “Lost-in-the-Middle” degradation, whereas discarding it forces redundant re-analysis (Lian et al., 13 Apr 2026). Its Dynamic Reasoning Context combines a sliding window of detailed reasoning with compressed Reasoning Digests, yielding

$$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$

On SWE-Bench-Verified, SWE-AGILE reports 24.1% success with the Qwen3-8B model using 2.2k training trajectories and 896 tasks, reduces redundant per-step reasoning tokens by ~28%, and uses compression-aware RL to reduce digest length by >30% per step without harming task success (Lian et al., 13 Apr 2026). This suggests that sensibility is increasingly treated as an issue of memory discipline and trajectory management, not only of raw inference quality.

3. Persistent identity, memory, and lifelong evolution

Synergy gives perhaps the most explicit architectural account of a sensible agent as a persistent computational subject. The paper defines an Agentic Citizen by

$$\mathrm{AC}(a, \mathcal{W}) \iff \mathcal{C}(a, \mathcal{W}) \land \mathcal{I}(a) \land \mathcal{E}(a),$$

where $\mathcal{C}$ denotes collaboration, $\mathcal{I}$ identity/personhood, and $\mathcal{E}$ evolution (Nie et al., 30 Mar 2026).

Its execution unit is the Session, described as the primary execution capsule beyond a simple chat transcript, and its runtime is the Cortex, which manages child sessions, delegation, and planning state via session-local DAGs (Nie et al., 30 Mar 2026). Sensibility is anchored in a multilayered identity substrate: a durable Holos profile, typed long-term memory with types {self, user, relationship, preference, asset, insight, knowledge, general}, writable notes, reusable skills, durable contacts/social relationships, an Agenda/TIME layer for obligations and events, and a distinct Experience Store (Nie et al., 30 Mar 2026).

The experience-centered learning mechanism is formally separated from semantic or procedural memory. Experience records are represented as

$e_i = (z_i, s_i, d_i, \mathbf{Q}_i),$

with inferred intent $z_i$, distilled execution script $s_i$, raw trajectory digest $h_t$0, and learned reuse value vector $h_t$1 (Nie et al., 30 Mar 2026). Reward assignment is vector-valued,

$h_t$2

tracking outcome, intent understanding, execution quality, orchestration quality, and expression quality, while delayed credit assignment updates reuse values for recalled experiences (Nie et al., 30 Mar 2026). The architecture therefore treats sensible behavior as inseparable from social continuity, typed autobiographical memory, and the ability to improve at inference time by recalling rewarded prior trajectories.

A plausible implication is that this line of work replaces the resettable function-call view of agents with a durable-agent view in which memory typing, social relationships, and temporal commitments are first-class runtime objects.

4. Situatedness, partial observability, and secure information flow

Bentley, Lim, and Ishikawa’s aspective agentic AI (A2AI) places sensibility in environmental situatedness and policy-constrained perception rather than in orchestration (Bentley et al., 3 Sep 2025). The framework is explicitly bottom-up and environment-centric, drawing on subsumption architecture and the biological concept of umwelt. Its three stated principles are Situatedness, Aspect Separation, and Reactivity and Asynchrony (Bentley et al., 3 Sep 2025).

The key object is the aspect, a filtered, policy-defined view of environment $h_t$3 for a given agent class:

$h_t$4

Here $h_t$5 is the aspect-generation function and $h_t$6 is the policy determining visible information (Bentley et al., 3 Sep 2025). In implementation, a perception agent ($h_t$7-agent) generates the aspect, and action agents ($h_t$8-agents) operate only within that aspect. Requests to change the environment pass through policy checks, and no single agent has global visibility (Bentley et al., 3 Sep 2025).

This architectural choice is evaluated against a typical shared-memory or group-chat style multi-agent architecture. In the information-breach experiment, A2AI reports zero leakage in all trials (30 runs), whereas the comparison architecture shows 37%–83% leakage rates depending on adversarial prompt type (Bentley et al., 3 Sep 2025). In the dynamic information-change experiment, A2AI correctly permits a medical-aspect update of an incubation period from 7 to 10 days while refusing the same request from the general-public aspect; the comparison system processes updates inconsistently and sometimes leaks them into public summaries (Bentley et al., 3 Sep 2025).

The paper’s claim that typical architectures leak up to 83% of the time and that A2AI enables zero information leakage positions sensibility as a property of architecture rather than prompt compliance (Bentley et al., 3 Sep 2025). This directly counters a common assumption in LLM-agent practice that information control can be delegated to prompting alone.

5. Human-facing sensibility: emotion, assistance, and unobtrusive interaction

In dialogue generation, sensibility is framed as the compatibility between emotional dynamics and external knowledge. The SEEK method for empathetic dialogue generation introduces Serial Encoding and Emotion-Knowledge interaction, using fine-grained utterance-level encoding to capture emotion flow and a dedicated module to resolve conflicts between candidate commonsense knowledge and current emotional context (Wang et al., 2022). Emotion prediction is modeled as

$h_t$9

and knowledge selection uses a gating score

$$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$0

so that only emotionally congruent knowledge guides response generation (Wang et al., 2022). On EmpatheticDialogues, SEEK is reported to outperform strong baselines on BLEU, Distinct-n, PPL, and a sensibility metric, with fewer cases of emotion-knowledge contradiction (Wang et al., 2022). In this setting, a sensible agent is one that does not merely retrieve relevant knowledge, but filters knowledge through affective coherence.

In XR, Sensible Agent operationalizes sensibility as low-friction assistance in socially and cognitively constrained contexts. The framework contains two coupled modules: an Action Recommendation Module for deciding what proactive assistance to provide, and an Interaction Adaption Module for deciding how to present it and receive input (Lee et al., 11 Sep 2025). Inputs include egocentric camera data, audio sensing, hand and head tracking, and contextual variables such as urgency, familiarity, social setting, and situational impairment. The system uses Large Multimodal Models, few-shot prompting, and chain-of-thought prompting to choose among multi-choice, binary confirmation, and icon-based cue formats, as well as among audio, visual, and audio-visual output modalities (Lee et al., 11 Sep 2025).

The design process included an expert workshop (n=12) and a data annotation study (n=40, 960 entries), and the prototype was evaluated in a user study (n=10) across AR and VR scenarios (Lee et al., 11 Sep 2025). Compared with a voice-query baseline, Sensible Agent yields NASA-TLX Raw 20.6 vs 43.3, with significance at $$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$1, SUS 81.3 vs 76.7, and preference 6.0/7 vs 3.8/7, significant at $$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$2 (Lee et al., 11 Sep 2025). Interaction time is longer—28.5s vs 16.4s—because of the added suggestion-confirmation step, while average end-to-end latency from context construction to prompt is ~6.2 seconds (Lee et al., 11 Sep 2025). These results support a narrower but important definition of sensible agency: an agent that reduces perceived interaction effort even when it adds a deliberate confirmation stage.

6. Utility shaping, score awareness, and broader conceptual implications

SAI shows that sensibility can also be defined at the level of utility shaping and state evaluation. Instead of predicting win probability only for a fixed komi, SAI models winrate for all komi values using a two-parameter sigmoid,

$$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$3

where $$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$4 estimates score difference and $$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$5 represents decidedness (Morandin et al., 2018). Training uses self-play games that occasionally branch with changed komi when a position is uneven, allowing the network to learn these two continuous parameters from varied game outcomes (Morandin et al., 2018).

The framework further introduces a parameter $$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$6 that biases evaluation away from pure minimal-win behavior. Rather than using only the winrate at the current komi, SAI can integrate winrate over a virtual-komi interval,

$$C_t = \left[ \bigcup_{i=0}^{t-N-1} (o_i, d_i, a_i) \right] \oplus \left[ \bigcup_{j=t-N}^{t-1} (o_j, r_j, d_j, a_j) \right] \oplus o_t .$$7

so that the agent can prefer larger wins or behave more naturally under handicap conditions (Morandin et al., 2018). On 7x7 Go, the method is reported to obtain very strong playing agents, while also giving a usable estimate of score difference and of how much the game is decided (Morandin et al., 2018).

Taken together, the literature indicates that sensible-agent research is not reducible to one paradigm. In some settings, sensibility is cognitive structure; in others, persistent personhood; in others, partial observability and policy isolation; in others, emotion-knowledge compatibility, interaction unobtrusiveness, or score-aware utility shaping. A plausible synthesis is that the term consistently marks a departure from stateless, globally exposed, purely reactive, or purely prompt-compliant agents toward agents whose competence depends on explicit architectural constraints, memory organization, and adaptation to the environment in which they are embedded.