cs.CLNov 3, 2026

From Autonomous Agents to Cognitive Symbionts: Toward AI-Augmented Human Intelligence

Authors: Hao Fei

Organizations: University of Oxford

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

For a decade, progress in artificial intelligence has been measured by how much a machine can do on its own. Large foundation models now write code and run multistep tasks, and personal agents act on a person's behalf. Yet people still forget what they once knew, miss what they are not looking at, fail to find the right words, and face hard choices without the benefit of their own history, because the machine that could help stands outside the situation. As AI moves into the physical world, attention has settled on one kind of intelligent individual, the embodied robot; we argue that the era of Physical AI holds a second kind of equal importance, the person augmented by AI. Alongside artificial general intelligence, which asks how intelligent a machine can be, stands AI-augmented human intelligence (AHI), which asks how much intelligence a machine can enable in a particular person. Realizing it calls for a new kind of system, a cognitive symbiont: a lifelong, embedded, real-time, private, low-friction, and proactive companion that senses what the person senses through external devices, keeps their context across years, models their situation and their mind, takes part in their thinking at the right moment, carries out everyday tasks with their context, and keeps learning the person as both change. We give the paradigm a conceptual core, the cognitive gap between what a person could bring to a situation and what they can hold in working memory, and organize its design space as a symbiosis loop of sensing, keeping, understanding, living alongside, acting for and thinking with the person, and growing, enclosed by a personal runtime and by trust. Because a symbiont changes what a person can do in front of others, it also raises risks of a new kind, from covert assistance in examinations to pressure on everyone to adopt it, and we treat their governance as part of the paradigm.

Figures & tables

Figure 2: The two halves of the endgame. AGI runs from chatbots to autonomous agents and is scored on what the AI completes. AHI runs from reactive assistants to cognitive symbionts and is scored on what the person can now do. Both share one base of models, agents, memory, and hardware, and both inherit every improvement to it. A system sits on one half or the other according to what it is evaluated on, so the same chat product is a chatbot when its answers are scored and a reactive assistant when what its user can then do is scored; today the first two stages of the AHI half are still scored on what the AI produces (Table 1 ). Arrows mark growth within a half, and the two halves are parallel, not successive stages.
Figure 3: (a) AI capability on isolated tasks keeps rising while coupling with a particular person stays nearly flat; the shaded region is augmentation that is never realized. The curves are a conceptual sketch, not measurements. (b) The mechanism of the shift: a causal loop in which stronger base models lower the cost of being understood, move interaction onto the body, and generate a new kind of context that makes the models more useful to each person in turn, without being used to train them.
Table 1: Four roles in human–AI relations, ordered by the relationship with the person. Down the table the AI knows more (a message, a task, your preferences, your life), moves closer (a chat box, a computer, the cloud, your body), and its value shifts from giving you information to making you stronger. A symbiont contains the first three roles; it does not replace them.
Figure 4: Two forms of intelligent individuals. Left: an AI-augmented individual, a person with a symbiont that senses and speaks through wearable devices. Middle: an AI as individual, an embodied robot. The two share one cognitive core, indicated by the common bar. Right, in lighter tone: devices in the environment that either form can reach into (Section 14.2 ). The paper’s scope is the left form. “AI as individual” is a functional description and carries no claim about consciousness or legal personhood.
Line of workRepresentativesWhat it has already settledWhat this paper asks nextTaken up in
1945 LifeloggingBush (1945) ; Gemmell et al. (2006) ; Hodges et al. (2006)recording a life, many views, memory supporthow a record turns into understanding of a person and timely help§ 6
1960 Human–computer symbiosisLicklider (1960)people and machines complementing each other in real time, thinking togetherwhat symbiosis looks like once machines understand language and images§ 8
1962 Augmenting intellectEngelbart (1962)the unit of augmentation is the person plus tools plus methodshow the research agenda changes once that unit is the object of evaluation§ 17
1991 Ubiquitous and wearable computingWeiser (1991) ; Starner et al. (1997) ; Mann (1997)computing woven into daily life; the first-person viewpoint of wearableswhat unnoticed computing can do once a model understands content§ 5
1998 Extended mindClark and Chalmers (1998)cognition can extend into tools outside the bodywhere the boundary lies for a tool that learns and takes initiative§ 4
1999 Mixed-initiative interactionHorvitz (1999) ; Amershi et al. (2019)principles for when to act and when to waithow these principles are rewritten for an all-day companion§ 8
Table 2: Lineage and nearest neighbors. Each line of work has settled something this paper relies on, and the fourth column states the question we take up from it.
Figure 5: The stack a symbiont is built on. From the bottom: wearable and embedded hardware, foundation models (one axis that runs from language to omni-modal to reasoning to world models), the personal runtime, the agentic harness, and the symbiont layer that this paper proposes. The side notes record what exists today and what the paper adds at each level.
Table 3: Five lines of technology, what each already gives a cognitive symbiont, and what it lacks. The last column maps every gap to the sections that address it; the gaps are the research space of Part II.
PropertyMeaningWhat breaks without itJarvisPersonal agents todaySections
Lifelongone relationship and one memory kept across yearsevery session starts from zero; the gap cannot close✓◐§ 6 § 10
Embeddedpresent in real activity, not waiting in a chat windowhelp can only happen afterwards✓✗§ 5 § 11
Real-timesteps in while the activity is still going and help is still usefulthe right words at the wrong time are no words✓✗§ 8
Privatewhat passes between the person and the symbiont stays theirs, and its delivery can be discreet where the setting allowsthe person cannot confide in it, and every cue disturbs the people nearby✓◐§ 5 § 8
Low-frictionnever asks you to stop, take out a phone, or explain the backgroundthe cost of use eats the benefit✓✗§ 8 § 11
Proactivecan judge whether to speak when you have not askedyou do not ask about what you do not know you are missing✓◐§ 8
Table 4: Six properties of a cognitive symbiont, and the Jarvis test. Each property is necessary; the third column says what fails without it. The properties pull against one another (proactive against low-friction, lifelong against private), and those tensions are the research problems of Sections 8 and 12 .
Figure 6: The symbiosis loop. Six stages run around the person; the runtime beneath them makes the loop run across devices and time, and trust encloses everything. The person’s own words enter the record directly. Silence is drawn as a formal output, and acting for and thinking with share one stage.
Figure 7: The device constellation. Each device carries a role on the way in (sensing, private input) or on the way out (private output, shared canvas), and often both; a runtime manages them as a single set of senses. The symbiont itself has no body and borrows every one of them. Neural interfaces and the devices of the environment are the far end of the constellation.
Table 5: Signal sources and what each reveals about the physical and the mental side of a life. “Omni” means the system can take whatever signal arrives; it does not mean every signal must be collected. The two sides of the table anticipate Section 7.1 ; the last column is the horizon on which each source becomes usable.
Table 6: Channels out, and when each fits. The private channels (first three rows) deliver help to the person alone, discreetly where the setting allows; the shared canvas (last row) lets the person and the symbiont, or the person and other people, look at the same picture. Jarvis’s holographic table belongs to the last row.
Figure 8: Three timescales of one record. Live context is large and mostly discarded, working state holds projects and open threads for days to weeks, and knowledge is small, curated, and persistent. Consolidation moves material up; recall moves the right slice down into the working context ctc_{t} of Definition 4.1 .
Table 7: Sources of personal context, and what each sees on its own. A chatbot knows a questioner, a coding assistant knows a programmer, a desktop agent knows an office worker, a band knows a body with a heart rate. Only one entry point that joins them can answer questions that cross the boundaries. Each source is authorized separately (Section 12.2 ).
Figure 9: One life, many views. The same handful of events (a meeting, a family plan, a bad night’s sleep) appears in six views, by logic, domain, person, lifetime, time, and nature. The views are projections of one record and can be switched at will.
Table 8: Views of the record and the questions they answer. Views are projections of one store: the record underneath is one, and a view is generated by the question being asked.
Figure 10: Two roads to the mind. Observation infers patterns from behavior and physiology; self-report records what the person says about intent, feeling, and reasons. Both feed one layered model, with every entry tagged by source and uncertainty; disagreements are marked, down-weighted, and resolved by asking.
Table 9: Layers of the personal model. Each layer carries its evidence and its rate of change, and the last column names the misuse it invites. The table is indexed by context: the identity, relations, and help-preference layers take different values for the self at work, the self at home, and the self speaking a home dialect, while experience and knowledge are shared. The last layer parameterizes the intervention policy of Section 8.4 .
Figure 11: Two axes of the backbone. Columns: the family of foundation models, from language to sight and sound to dynamics, each extending what counts as evidence. Rows: how much system is organized around the model, from a single pass to a process inside an agentic harness. The two axes are independent, since a harness can be wrapped around any of the three models; the dashed path is the likely order in which systems will move through the grid, and the heavy line is the information interface of the personal model, which is the same in every cell.
Figure 12: Cognitive full-duplex. One exchange in a team meeting: a colleague raises a figure the person reported last quarter, the slow loop retrieves and checks it, the fast loop places a one-line cue in the glasses within a second, the person answers, and the cue is withdrawn once it expires. The timeline below follows the exchange across the four tracks of the external world, outward expression, the private channel, and the symbiont’s internal state. The help is a reminder of the person’s own figure, the kind of assistance most settings accept (Section 15.5 ), and most of the time the right output is silence.
Table 10: The three decisions of the intervention policy. Silence is always one of the answers to how , and the last column makes it the default.
Figure 13: The ladder of assistance. Tier 1 is where today’s agents stop; a symbiont does the same tasks with the person’s context in hand. Tier 2 is thinking and feeling with the person, and Tier 3 is living among others. The depth of the personal model required grows with each tier, and caring runs across all three.
Table 11: The same task, done by an agent and by a symbiont. The task is unchanged; what changes is how much must be asked, how the task is framed, and how often the result has to be corrected. The last column makes the difference measurable.
Figure 14: The symbiont handshake. Four rings of authorization around each person, from visible to committable; exchange happens only within the ring both have opened, and the inner model never crosses.
Table 12: A catalogue of cognitive-level functions as tests of the framework. Every row names the layers of the personal model it needs, the channel it uses, its timescale, the failure that would best discriminate a good system from a bad one, and its maturity. Section 13 maps rows to populations and Section 17 to evaluation.
Table 13: Seven places where a symbiont can learn. Most personalization lives in the first three rows and already creates value; the fourth and fifth are self-evolution; the sixth is hardware; the last is the paper’s own object of evaluation. Sounding more like you is not the same as understanding you better.
Figure 15: The learning spectrum. The lower band runs from changing only memory to changing all parameters; each mechanism is rated for the data it needs, its stability, and its reversibility, and the paper argues for the middle-left region. The upper band is harness-level learning, which runs in parallel, stays readable and reversible, and is where self-improvement should mostly happen.
Figure 16: Co-evolution of a person and a stack. Above the axis, events in the person’s life; below it, changes in the system: base model upgrades, routines and skills the harness accumulates, new hardware. Dashed links show life events triggering system learning and system changes altering what is observed next. The heavy line is the record and personal model, the one thing never replaced. In the fiction, the counterpart of this growth is Jarvis’s cognitive core eventually gaining a body as the Vision (Section 14.2 ), and the counterexample is Ultron, a separate project whose goals diverged from its maker’s (Section 15.7 ).
Figure 17: The personal runtime. Base capabilities can be swapped freely; the record and personal model at the top are never swapped. The harness gets things done; the runtime puts them on the right device, with the right permission, at the right time. The trust layer runs through everything.
Figure 18: The trust curve. Data authorization (solid, read on the left axis) and action permission (dashed, read on the right axis) climb in steps as time together and delivered value accumulate. Each data level names what must be proved before the next is opened, each action level is earned a little after the data level it needs, and a breach sends authorization back one step. The system has to be useful at the lowest level already.
Table 14: Design principles for cognitive symbionts. Each is derived from one stage of the loop, states a requirement on the system, and names a check. They are design positions, not an interface standard.
Figure 19: A map of beneficiaries. Horizontal: how much of a population’s gap a symbiont can close today (observability times closability). Vertical: the size and stakes of the gap. Bubble size is population; color is the maturity of the evidence. Positions are the author’s judgments, not measurements.
Table 15: Who gains what. Each row summarizes one population of Section 13 ; the functions column points into Table 12 , and the last column locates the population in Figure 19 .
Figure 20: Three directions of harm around an augmented person, and their sum at the scale of a society. Harm to the person comes from the system, its provider, or an attacker; harm through the person reaches the people and institutions the person deals with; harm from institutions comes from those that deploy the paradigm for their own ends. Chips name the risks, and section marks point to where each is treated.
Table 16: Contested uses of the symbiont’s own functions, drawn from Table 12 and Section 15.5 . For each function: what it is for, the misuse it invites, the settings in which it is contested, and the default the paper proposes.
Table 17: Dimensions of evaluation. Each observable is paired with what it cannot establish alone; no single dimension is sufficient.
 ProblemWhat has to be worked outDepends on
A Modeling a person and their record
A1Context and knowledge modeling at scale ★how a life’s multimodal stream becomes a record that can be queried in milliseconds, revised, and viewed in many ways; what is kept and at what fidelity; how consolidation and recall across the three timescales are automated§ 6.1 § 6.3
A2Grounded modeling of the mindhow the observation road and the self-report road are fused, and how their disagreements are represented and resolved§ 7.2
A3Many selveshow context-indexed personas are represented and triggered, and how they are kept from collapsing into one average person§ 7.3
A4Joint physical and mental world modelslearned from first-person streams, coupled in both directions, and able to detect their own staleness§ 7.5
A5Backbone-agnostic persistencehow the information-level interface is specified, and how a migration across backbones is verified to have preserved the relationship§ 7.4 § 10.4
Table 18: Open problems of the paradigm, in six groups. Each problem names what has to be worked out and the sections it depends on; the starred problems are the ones on which the first usable systems most depend.
Figure 21: From components to a life. Six stages of accumulating capability, the evidence that would justify moving to the next, and the open problems of Table 18 mapped onto them.
More in this topic

Human-AI Collaboration

Human-AI collaboration studies how people and AI systems divide, coordinate, and adapt work to support decisions, creativity, and learning.

See all 458 papers

Explore similar work

CardsList
Aug 11, 2026

ComBodied Agents: a New Paradigm of Human-Centric Agentic AI

After an older adult misses a medication dose, a software agent can send another reminder and an embodied agent can bring the medication. Yet neither explains whether the person forgot, is confused, has side effects, or deliberately refused, nor what support is appropriate. This reveals a structural gap in Agentic AI: Digital Agents primarily transform software states, while Embodied Agents transform physical states; neither makes a person's evolving state and agency the primary object of modeling, intervention, and evaluation. We introduce Combodied Agents, a human-centered paradigm that perceives, models, predicts, and supports individual human-state trajectories over time, using software tools, sensors, wearables, robots, and human services as action channels rather than end goals. We unify fragmented capabilities across personal assistants, health agents, AI companions, and adaptive human--AI systems into a closed loop: event-based multimodal perception reconstructs meaningful personal events; longitudinal, correctable memory provides temporal context; Personal World Models estimate future personal states and outcomes under alternative decisions and interventions; and an admissible intervention policy selects proportionate support under consent, uncertainty, safety, reversibility, and user control. Feedback from the person and environment updates the loop. Rather than requiring an exhaustive Human Digital Twin, the framework uses purpose-bounded, uncertainty-aware, user-correctable representations. We organize the design space by human-state targets, relational contexts, and agent roles, and propose scenario-centered evaluation, agency-preservation metrics, benchmark requirements, edge-native personal models, and governance directions. Combodied Agents shift Agentic AI from external task completion toward sustained human benefit.

Qianggang Ding, Xingyao Wang, Rui Feng +19

Universit´e de Montr´eal, Canada · Mila – Quebec Artificial Intelligence Institute, Canada · Institute of Advanced Intelligence and Computing (IAIC), A*STAR, Singapore +13

Human-Centered AIAgentic AI
AskPDF
Sep 7, 2026

When Intelligence Becomes Agency: A Theory of Governed, Proactive Agency for Symbiotic AI Systems

Persistent AI assistants are intended to extend human attention, memory, and coordination across changing digital and physical environments. To be truly useful they must do more than just act when asked. They must decide on their own whether a situation warrants behavior at all, when it does and in what mode, whether to act, ask, monitor, defer or deliberately refrain. We call this the activation problem. Research on commitment, appraisal, mixed-initiative interaction and delegation each illuminates part of it, but none ties situated activation to continuing authorization and accountability. This paper develops a conceptual and formal framework for governed proactive agency, organizing behavior across time through perception, intent, affective-conative appraisal, constraint, and feedback. It distinguishes autonomous and delegated agency and defines symbiotic agency as delegation under a standing, revocable mandate, with continuing coupling to the principal's situation, calibrated inference of their condition, and bounded personalization. The distinctive contribution is an integrated account linking activation decisions to authorized perception, behavior selection, authority containment, traceable restraint, and constrained adaptation, with behavioral episodes as the unit of analysis. Through an agency classification method, an evaluation framework, proposed benchmark scenarios, and a reference architecture, the account provides a basis for specifying and assessing whether assistance is warranted, timely, authorized, and answerable beyond task completion alone. It is intended to guide the development and evaluation of always-present personal assistants and embodied support systems that augment human capabilities while preserving the principal's authority and judgment.

João Dias Ferreira

Wyrde AI

Agent EvaluationProactive AssistanceAI Agent GovernanceHuman Oversight of AIHuman-AI Collaboration
AskPDF
Aug 24, 2026

AI Agents Push Humans Out of the Loop

AI agents pose significant risks as they are granted increasing autonomy. A commonly proposed solution is human oversight and keeping a ''human in the loop'', but this is not a simple solution: Not only do current approaches to AI agent design impede effective human oversight, but the cognitive capacities required for it are also themselves degraded by extended use of AI systems. This position paper argues that current approaches to the development and deployment of AI agent systems do not support effective human oversight -- they contribute to its degradation. To address this, a top priority in the advancement of AI agents should be supporting the situated goals and cognitive requirements of effective human oversight, treating the human needs of overseers at the same level of importance as AI agent capability. To put this idea into practice, we connect work on automation and human-computer interaction to AI agent processes, outlining design-level affordances and organizational protocols that (1) support overseers in exercising critical judgement and (2) counteract the skill atrophy that arises from extended use of automation. We urge developers and deployers to adopt these or similar approaches. Without explicit support for the cognitive demands of effective human-agent interaction, AI agent systems will continue to passively incentivize the degradation of the very human skills they rely on.

Margaret Mitchell, Avijit Ghosh, Samir Passi

Hugging Face · Data & Society

AI Agent SafetyHuman-in-the-Loop AIHuman-Centered AIHuman Oversight of AI
AskPDF