News / Research watch · Agent infrastructure
The Survey That Tells Agents to Find Their Own Work: Google-UIUC's 203-Page Proactive-AI Map
Agents can read code, call tools and run long tasks — but in most stacks, noticing what needs doing and deciding it is worth doing still belongs to the human. A 23-author survey from UIUC and Google, "Proactive AI for Superintelligent Agents", consolidates 1,410 cited works into a map of that missing layer: proactiveness, the capacity to start purposeful work without an explicit request. We verified the artifact directly — the PDF is exactly 203 pages by file metadata, the author list and 2026 manuscript citation check out on Google Research's publications page, and the MIT-licensed companion repo organizes the full 1,410-work library, 80 representative works across 13 domains. For builders running long agents on metered APIs, the load-bearing part is one chapter: Arbitration — the four gates an agent should pass before it spends your tokens on work nobody asked for.
Verified directly: the survey exists and checks out — 23 authors led by Ke Yang and Jiateng Liu (UIUC) with Google co-authors (Zifeng Wang, Chen-Yu Lee), cited as yang2026proactiveai on Google Research's publication page; the hosted PDF is 203 pages; the companion repo (EmpathYang/Proactive-AI-for-Superintelligent-Agents, MIT) organizes 1,410 cited works — 80 representative ones across 13 domains — with the library updated October 2026.
The framework: six capacities — Awareness, Anticipation, Agenda, Arbitration, Action, Adaptation. The hinge is Arbitration: candidate interventions pass gates of uncertainty (do I actually understand the need), utility (does the gain beat the cost), timing (now or later), and risk/reversibility (how bad if wrong). Evaluation runs at three levels — task, mechanism, user — with supervision burden and the acceptance-rate trap called out explicitly. A cross-domain bottleneck: feedback speed, cost and reliability cap how fast an agent can safely iterate.
Not established: no peer-reviewed venue yet (a 2026 manuscript hosted by Google Research); it is a survey, not a system — no benchmarks, no products; no China-vendor content — the China connection runs through this site's own coverage line, where DeepSeek's dsh-superagent and Tsinghua/Zhipu's HarnessPro already practice pieces of the same cost-discipline problem.
The facts, graded (Nov 7)
- The artifact, verified by direct read. PDF hosted on Google Cloud storage (pubtools/1048480): 203 pages confirmed from file metadata. Authors: 23, anchored by UIUC (Ke Yang, Jiateng Liu, Hanghang Tong, Jingrui He, Chenyan Xiong, Jiawei Han, ChengXiang Zhai et al.) plus Google co-authors (Zifeng Wang, Chen-Yu Lee). Cited as
@misc{yang2026proactiveai}, 2026; no conference venue attached. - The library, verified in the repo. GitHub
EmpathYang/Proactive-AI-for-Superintelligent-Agents(MIT, 7 commits, 18 stars): 1,410 cited works organized under the framework, 80 representative works analyzed, library data updated October 2026. The README states the survey covers works from 1937–2026. - The six capacities (6A). Awareness (maintaining the situation), Anticipation (context → unspoken needs), Agenda (prioritized persistent goals), Arbitration (gating initiative via uncertainty, utility, timing, risk, reversibility), Action (intervening within authority), Adaptation (calibrating future initiative from outcomes and feedback — accept, reject, modify, ignore, withdraw, takeover each read in context).
- The 13 domains. Coding, GUI agents, information access, embodied & robotics, security, AutoResearch, smart home, smart city, healthcare, law, education, entertainment & media, agriculture — each mapped on a five-rung initiative ladder from reactive to agenda-generative.
- The evaluation stack. Task level (did outcomes improve; did saved time beat added cost), mechanism level (situation read, anticipation, timing), user level (supervision and correction burden, control, trust matching reliability). The acceptance-rate trap: an agent can learn to package suggestions better without getting better at judging worth.
- The bottleneck finding. Across domains, what limits proactive iteration is feedback: CI and compilers close in seconds or minutes; healthcare, agriculture and wet-lab feedback takes a treatment cycle, a season or an expensive experiment. Feedback speed, cost and reliability set the safe tempo of initiative.
- What remains open. No venue/peer review, no benchmark suite, no reference implementation. The 203-page figure we confirmed from the PDF itself; media summaries (NLPer, Nov 5) relayed the framework consistently with the primary sources.
Verification on Nov 7, 2026: Google Research publication page, GitHub repo README and project page read directly; PDF page count read from the file itself (9.1 MB, /Count 203). Framework claims cross-checked against the repo README structure. No independent assessment of the survey's scholarly quality was attempted.
Arbitration, in builder terms: the four gates before your tokens move
The survey's most load-bearing idea for anyone running agents is that acting and not-acting must be compared before acting. A scheduling conflict is just data until the agent weighs it against route and timing, estimates whether preparing a remote-call option actually helps, and only then earns a reason to speak. Translated to an agent stack on metered APIs, the four gates are an engineering checklist: uncertainty — would a human call this a real need or a thin clue (isolated validation before proposing a patch); utility — expected gain versus compute, time, interruption and correction cost; timing — hold the finding for a natural task gap; risk/reversibility — database writes, merges and outbound messages need prior authorization precisely because they are hard to undo.
The survey's failure gallery is familiar to anyone who has run a proactive assistant: every suggestion is individually reasonable and collectively unsurvivable — refactor here, docs there, upgrade this dependency — while the user is shipping. The remedy it documents is structural, not behavioral: gate candidates before execution, keep discoveries parked until the timing gate opens, and treat "do not disturb" as a designed outcome rather than a failure.
The evaluation chapter lands on this site's weekly arithmetic: task-level gains must net out supervision burden at user level. An agent whose task list grows while the human's cleanup queue grows faster is negative-value regardless of its benchmark scores.
Context: the Chinese-lab line has been practicing pieces of this
The survey has no China-vendor content, but the problems it systematizes are live on this site's beat. DeepSeek's open-sourced dsh-superagent (full translation) engineered exactly the Action-and-cost half — batching, event filtering, speculative execution — for 24/7 persistence. The Tsinghua/Zhipu HarnessPro paper made arbitration's utility gate quantitative: +2.0pp pass rate at −26.1% per-task cost by choosing reasoning intensity and stopping points per task. What the survey adds is the frame around those point solutions: Awareness and Anticipation before the spend, Arbitration before the spend is justified, Adaptation after it.
For builders, the cost arithmetic is the same one we run on model pricing every week: a proactive agent's "worth-it" gate is the per-task cost question — DeepSeek V4.1-Flash off-peak at $0.22/1M versus GLM-5.3 at $1.40/$4.40 decides which behaviors your stack can afford to let an agent initiate unprompted, and the comparison table holds the ladder. Authority, as the survey frames it, is granted on earned trust: start agents on reversible, sandboxed initiative; expand scope only as the adaptation loop earns it.
Watch for: the survey picking up a venue and citations, the repo library growing past 1,410, and any vendor shipping arbitration-gated proactivity as a product feature — that last one would move this from research watch to launch coverage.
Access from outside China — and the usual disclaimer
This site tracks Chinese model APIs for international builders: official endpoints, pricing and open-weights status, plus the agent-infrastructure work built on them. The models underneath the cost arithmetic live in the comparison table and the cheapest-API guide; the agent line is covered in the dsh-superagent and HarnessPro briefs.
Independence disclaimer: ChinaModelAPI is an independent information site, not affiliated with the survey's authors, UIUC, Google or any model vendor. Artifact facts are from direct reads (publication page, repo, PDF metadata); framework summaries attribute the survey; nothing here is a routing, availability or pricing promise, and no assessment of the survey's scholarly merit is implied.
Primary sources
- Google Research — publication page (23 authors, abstract, 2026; read directly)
- Survey PDF (pubtools/1048480) — 203 pages confirmed from file metadata; read directly
- GitHub — EmpathYang/Proactive-AI-for-Superintelligent-Agents (MIT; 1,410 works, 80 representative, 13 domains; README read directly)
- Project page — interactive overview; library data updated October 2026
- NLPer / AINLP (WeChat, Nov 5, 2026) — the walkthrough this brief cross-checked; consistently relayed
- 站内对照：DeepSeek dsh-superagent 开源 · HarnessPro 成本纪律论文 · 最低价指南 · 对比表
FAQ (2026)
What is the Proactive AI for Superintelligent Agents survey?
A 203-page survey by 23 authors from UIUC and Google (Ke Yang, Jiateng Liu et al.; cited as yang2026proactiveai, hosted on Google Research publications), consolidating 1,410 cited works — 80 representative ones across 13 application domains — on how agents can find and start worthwhile work without an explicit request. The companion GitHub repo (MIT) holds the paper and the structured library, updated October 2026.
What is proactiveness, in the survey's definition?
The capacity to initiate purposeful action without an explicit request — exercising the discretionary room that instructions leave open (trade-offs, gaps, contingencies, replanning). The survey treats it as a third axis beside intelligence (capability) and alignment (serving human values), aiming for calibrated initiative rather than over- or under-acting.
What are the six capacities (6A framework)?
Awareness (maintaining the situation), Anticipation (turning context into unspoken needs), Agenda (organized, prioritized goals), Arbitration (deciding whether to intervene now), Action (intervening within granted authority), Adaptation (learning from outcomes and feedback).
What are the arbitration gates?
Before intervening, a candidate action passes four checks: confidence (does it really understand the need, or is it guessing from a thin clue), utility (does the expected gain beat the compute, time, interruption and correction cost), timing (now, or after the user finishes the current step), and risk/reversibility (if it is wrong, how bad and how recoverable). Waiting and holding a finding for a natural gap are part of the design, not failures.
How should proactive agents be evaluated?
Three levels: task (did the outcome improve — did saved time beat added cost), mechanism (did it read the situation, anticipate, choose timing well), and user (supervision and correction burden, experience, control, and trust matching actual reliability). The survey warns that optimizing suggestion-acceptance rate alone can teach an agent to package suggestions better without judging worth better.
Why does this matter for API builders?
Long-running agents on metered APIs live inside the arbitration problem: every proactive step is tokens billed against uncertainty. The survey's gates and three-level evaluation are a checklist for deciding which proactive behaviors your stack can afford — and its cross-domain finding is that feedback speed, cost and reliability set how fast an agent can safely iterate.