Proactive AI's 6A Framework for Superintelligent Agents: Joint Survey

Background: From Reactive to Proactive Agents

As early as 1995, Wooldridge et al. defined four core agent properties: reactivity, social ability, autonomy, and proactiveness. While reactivity, social ability, and autonomy have seen substantial progress with large language models and alignment research, proactiveness lacks a systematic research framework. This gap is especially critical for long-horizon tasks where humans cannot pre-specify every step or foresee all contingencies. Proactiveness enables agents to clarify ambiguous instructions, explore alternatives, and adjust plans without waiting for the next command.

Joint Survey and the 6A Framework

Researchers from UIUC, Google, Stanford, UC Berkeley, and CMU present the survey "Proactive AI for Superintelligent Agents" (project page:

https://empathyang.github.io/Proactive-AI-for-Superintelligent-Agents/

; Google Research:

https://research.google/pubs/proactive-ai-for-superintelligent-agents/

). The survey integrates insights from AI agents, embodied AI, information retrieval, computational biology, and economics/finance, citing over 1,400 references with 80 representative works analyzed in detail.

The authors propose the 6A mechanism framework — six interacting capabilities that support every proactive action:

Awareness — continuously track multimodal state (dialogue, calendar, device status, preferences), maintain hierarchical memory, and actively acquire missing information (e.g., a medical diagnosis agent asking targeted follow-up questions).

Anticipation — predict upcoming human needs and environmental changes using user models and world models; compare counterfactual outcomes of acting vs. waiting (e.g., an educational agent deciding whether to intervene when a student lingers on a problem).

Agenda — convert anticipated needs into committed goals, manage cross-session task lists, handle priority conflicts, and preserve rationale for later resumption.

Arbitration — decide whether to intervene now based on evidence reliability, expected benefit vs. cost, timing appropriateness, and risk/reversibility (e.g., a security agent weighing immediate account lockdown vs. further investigation).

Action — choose among reminders, questions, drafts, pre-filled content, or direct execution; provide previews, confirmations, undo, and handoff mechanisms.

Adaptation — interpret feedback (tool results, human acceptance/modification/override) for rapid in-task correction and long-term calibration of preferences, thresholds, and collaboration styles.

Distinguishing Proactiveness from Related Concepts

The survey separates intelligence level (understanding/reasoning), alignment (conformance to human intent), proactiveness (self-initiated valuable actions), and autonomy degree (authorized independent execution range). A household robot example illustrates: it may proactively suggest restocking detergent (proactiveness) but leave payment approval to humans (autonomy degree).

Historical Research Lineages

Proactive AI draws on pre-LLM research: control theory (feedback adjustment), planning/world models (future state prediction), agent theory (goal formation and commitment), context-aware computing (situational judgment), and HCI (interruption, mixed-initiative interaction). LLMs and agent harnesses (memory, tools, skills, execution management) now expand the feasible task scope.

Evaluation Methodology for Proactive AI

Standard agent benchmarks give a goal and measure completion. Proactive evaluation must retain unspecified needs and assess task discovery accuracy, missed opportunities, and spurious interventions. The survey structures evaluation at three layers:

Task layer — outcome quality, time, compute, and human effort.

Mechanism layer — diagnose failures in awareness, anticipation, agenda, arbitration.

User layer — workload, sense of control, trust, and disruption.

A table-clearing robot example: task layer checks if the table is correctly cleared and resources used; mechanism layer verifies correct detection of utensil usage and timing; user layer records confirmation/correction time and trust.

13 Application Domains

The survey covers coding, GUI operation, information retrieval, embodied robotics, security, automated research, smart home, smart city, healthcare, legal, education, entertainment/media, and agriculture. Each domain differs in evidence availability, permissible actions, and feedback latency.

Digital work — code changes in isolated branches, cross-app material preparation, proactive retrieval.

Safety, healthcare, legal — proactive expansion of investigation (vulnerability variants, new test results, contract clause ripple effects) with human gating for critical actions.

Automated research — joint assessment of novelty, feasibility, verifiability; reproducible experiment reports.

Physical domains — robot collaboration (intent signaling), agriculture (weather, crop stage, resource scheduling).

Education & entertainment — metrics beyond immediate completion (long-term retention, challenge preservation).

Maturity varies: deployed home systems mostly trigger/predict; automated research explores hypotheses under given directions; city systems often leave prediction to humans. Long-term responsibilities (cross-visit medical follow-up, whole-case legal management, longitudinal learning plans) remain largely open.

Three Cross-Domain Trends

Foundation models + domain-specific methods — similar reasoning models meet different execution/verification conditions.

Agents owning longer workflows — responsibility extends to end-to-end processes.

Reliable feedback & verification — continuous improvement depends on trustworthy learning signals.

Open Challenges Toward Superintelligent Agents

Mechanism-level evaluation & long-term testing — need diagnostics that pinpoint failures and experiments spanning weeks/months to catch preference drift, task interruption, and stale information.

Safety from perception onward — malicious web/email content may induce unauthorized tasks; continuous sensing may collect irrelevant private data. Systems must separate trusted instructions from external data, scope tasks/data, and allow human review/revocation.

Multi-stakeholder governance — in homes, organizations, public services: who proposes goals, who approves, who bears responsibility? Oversight interfaces must be understandable and avoid rubber-stamp fatigue.

Calibrated proactiveness — adjust proactive behavior based on evidence, timing, risk, preferences, and permissions; prepare/suggest/execute; revise/stop/handoff when conditions change. This judgment must be validated in long-term, auditable deployments.

Key References and Resources

Wooldridge, M., & Jennings, N. R. (1995). Intelligent Agents: Theory and Practice.

Proactive Agent (Tsinghua), ProAgent (CUHK), Meta Muse, OpenAI Dots, Anthropic Duet — early product/research explorations.

Survey paper: "Proactive AI for Superintelligent Agents" (UIUC, Google, Stanford, UC Berkeley, CMU).

Project page with 1,400+ citations:

https://empathyang.github.io/Proactive-AI-for-Superintelligent-Agents/

Google Research publication:

https://research.google/pubs/proactive-ai-for-superintelligent-agents/