Proactive AI for Superintelligent Agents
AI agents can reason, plan, use tools, and carry out long-horizon tasks.
Many still rely on a person to identify worthwhile work, and to start it.
1,410 works cited · 80 representative works analyzed in depth · 13 application domains
Why it matters
A capable agent can still depend on someone else to notice the work.
It can finish the assigned task while a person still has to notice what needs doing next.
Long-horizon work cannot be fully specified at the outset. Proactiveness reduces the burden of monitoring and re-instructing.
Why superintelligence needs it
Intelligence and alignment leave one capability unaccounted for.
- Intelligenceexpands what an agent can accomplish.
- Alignmentconcerns whether its behavior serves human intentions and values.
- Proactivenessadds the capacity to recognize worthwhile work and initiate it without an explicit request.
Reactive
The agent acts when asked.
Given a brief, a reactive agent works to fulfil it. A person has already identified the work to be done, and anything the brief leaves open waits for the next instruction.
Task execution space
An instruction specifies only part of the work.
The brief states a goal, constraints, resources and a plan: the explicit task instructions. What it leaves undecided is the residual decision space.
Definition
Initiative lives in what the instruction leaves open.
Proactiveness is the ability to exercise delegated residual discretion.
A proactive agent resolves trade-offs, fills procedural gaps, handles contingencies and replans, while respecting the task’s stated goals and constraints.
Spectrum · reactive
Reactive response
The agent follows step-by-step instructions. The human operates it, one command at a time. Reactivity and proactiveness are two ends of one spectrum, and most fielded systems sit between them.
Spectrum · mixed initiative
Mixed-initiative assistance
The agent proposes a plan and the human steers it. The human becomes a collaborator.
Spectrum · self-initiated
Self-initiated agency
The agent advances the work and reports progress. The human becomes a reviewer.
Lineage
Proactive agency renews a much older agenda.
Proactive agency draws on control theory, planning, agent theory, ubiquitous computing, mixed-initiative interaction and goal reasoning. Each tradition expanded the decisions an agent can make beyond a fully specified objective.
Lineage
Foundation models recombine these traditions.
They integrate the earlier mechanisms in open-ended, language-mediated settings. That makes proactivity more general, and calibrated proactivity more necessary: when should the agent move first, how far should it go, and when must it ask, defer, or stay silent?
Mechanism
Six capacities turn context into calibrated initiative.
Awareness maintains the situation and anticipation identifies future needs. Agenda formation turns them into commitments. Arbitration decides whether intervention is warranted. Action intervenes within authority. Adaptation learns from the result.
Anticipation and Arbitration address two distinctive demands of self-initiated work. The other four have established roles in general AI agents, and their proactive forms can require substantial changes.
The loop closes
Adaptation feeds every other capacity.
Fast reflection corrects a single intervention. Slow calibration adjusts how much initiative the agent should take next time, and with whom.
Applications · 13 domains
Initiative and authority grow on different axes.
Each bar spans where a domain’s agents get their tasks, from executing an instruction to pursuing a standing mandate. Its row shows what they are permitted to do on their first move, and its number counts the representative works the survey analyzes in depth. No domain has reached the fifth rung, where agents generate their own agenda.
Authority · gated action
Most domains act behind a gate.
Coding and GUI agents already work from forecasts and standing mandates, yet their consequential steps still pass review. Healthcare and law remain clinician- and attorney-gated.
Authority · suggestion
Three domains only suggest.
Information access, smart cities and education stay at suggestion, and entertainment spans suggestion and action. Their binding constraints differ: education and entertainment have slow or noisy feedback, smart cities have diffuse authority and public accountability, and for information access they are the user model and benchmark coverage of appropriate silence.
Authority · unsupervised
No domain acts unsupervised.
AutoResearch is the extreme case. Its systems can combine deep execution autonomy with task-scoped initiative under a human-specified top-level question, and they act inside a gate. Its binding constraints are novelty and significance, and the cost and latency of the physical oracle.
Across industries
Three patterns recur across domains.
Most systems pair a foundation-model core with domain-specific machinery. A growing subset executes extended workflows. Further improvement is constrained by verification and feedback infrastructure: how cheaply, quickly and reliably agents can learn from self-initiated actions.
Earned autonomy
Autonomy is the authority humans grant.
Intelligence provides the competence to reason and execute. Alignment directs it toward human intentions and values. Proactiveness enables its self-initiated use in the residual decision space. Exercised reliably, they establish a track record that can earn trust, and autonomy is granted on that demonstrated reliability.
What comes next
Toward calibrated proactiveness.
The survey closes with three directions: advancing capability and its evaluation; governing initiative with safety boundaries, consent and accountability; and application frontiers from personal agents to embodied and interconnected ecosystems.
Epilogue
This time, the agent speaks first.
It noticed a change, judged that acting was warranted, timely and acceptable, drafted a remedy within its authority, and left the decision with you.
