Whose Ally Is ‘Your’ Agent? ‘Computational Egonomics’ and the Allocation of Human Agency

Source: Michael Schrage; MIT Initiative on the Digital Economy working paper · 2026-08

Abstract. The dominant narrative about agentic AI impact keeps score in capability, atrophy and displacement. This paper argues that the consequential impact lies elsewhere: in the allocation of human agency. Every delegated action reinforces some human capacities and erodes others. Drawing on Thomas Schelling's classic ‘Egonomics or The Art of Self-Management,’ we model the self — not one decision-maker but several, present and future, sharing a body but not always a preference — as a portfolio of strategic players. We assert the Quantified Self movement failed because it measured the outputs of this game without ever instrumenting its players. We specify an alternative: a ‘Multiple Selves’ architecture in which commitment devices generate their own evaluation data, governance is intra-personal mechanism design, and the central claim is rendered measurable. The defining competition in agentic AI is not ‘model capability;’ it is the architecture of measurement and feedback wrapped around the human-agent interaction. There is no ‘neutral default.’ We recognize this approach has enormous implications for ‘behavioral economics’ and finance, as well as for organizational design.

Introduction. Past midnight before the earnings call: The CFO clears their queue of seventeen agent-generated recommendations — approve, approve, approve, decline — and goes to bed. The enterprise has just been governed. The question no dashboard asked was: just which ‘self’ did the governing? Not which person — the CFO’s approvals are not in doubt — but which version of them? The ‘strategic self’ that set the quarter’s priorities? Or the ‘exhausted self’ that wanted the queue empty? These are the human questions agentic AI forces and almost no enterprise system is instrumented to see. Bluntly: the most important impact of agentic AI will not be ‘task automation’ at scale; it will be the reallocation of human agency - not autonomy from the agent but the capacity to revise one's own commitments rather than simply act on them.

The techno-futurists keep score in capability and displacement. They debate whether ‘autonomous’ agents will book the travel, close the trades, write the code, run the supply chains, and how many knowledge workers will ‘retire’ as they do. This framing inverts consequence: Agentic systems do not merely anticipate, coordinate, and execute decisions; they’re alternately chartered and harnessed to allocate attention, initiative, memory, judgment, and deliberation across competing modes of human conduct. Every agentic interaction quietly reinforces some human capacities and lets others shrivel. Every delegation is also a choice about what kind of judgment gets strengthened, what kind gets weaker, and who the human becomes — for individuals and for teams.

However, the deeper and more consequential design question is not what can the agent do? but which human capabilities does the agent reinforce, undermine, or invite? An agent that removes friction from impulsive decisions cultivates a measurably different human future than one that slows the principal long enough to provoke or evoke - reflection. Agents engagingly optimized for ‘convenience’ can systematically weaken strategic judgment even as they measurably raise productivity. Agents built to elicit assumptions, commitments, and tradeoffs may feel slower at first while compounding human agency over time.

The crucial determinant is not agent capability; it is the intentional architecture of measurement and feedback wrapped around the interaction. That architecture has a name in waiting — the quantified selves, plural, governed, reflective — and a paradigm old enough to carry a few Nobel Prizes and new enough to be buildable only now: computational egonomics. Whether agentic, generative, and predictive AI capabilities are provisioned to serve that architecture - or automatically bypass it - is the intelligent 'intelligence' question, challenge and opportunity. Every organization seeking to productively deploy agentic AI at scale needs to understand that.

Personal informatics, commitment-device research, behavioral design, AI alignment, and people analytics each address important aspects of self-regulation and delegation. This contribution is different: an architecture that treats the ‘allocation of agency’ itself as the primary object of measurement and governance. Agentic AI exposes a hidden and flawed assumption embedded in classical principal-agent theory: that the principal is singular. Computational egonomics begins from the opposite premise: The fundamental challenge is not ensuring agents faithfully serves their principal; it is determining which principal those faithful agents serve. With apologies to Walt Whitman, principals are large and contain multitudes. When is ‘human agency’ more bug than feature – and vice versa?

Related work. Nobelist and strategist Thomas Schelling named the problem classical economics could not formalize. His 1978 ‘Egonomics, or the art of self-management’ article boldly argued that strategic interactions between a person's present self and the future selves who must live with the present self's The Quantified Self movement of the 2010s was a serious idea served by trivial instruments. Its Druckerian premise — that measurement enables management, that what is tracked can be improved — was correct and remains so. Its implementation was, no pun intended, bloodless: Step counts, sleep cycles, glucose curves, screen time, and heart-rate variability produced a torrent of data and a trickle of transformation. People came to know measurably more about themselves but few 'became' measurably more.

Even worse, it had no theory of purpose. Tracking assumed that more data plus a dashboard equals measurable and/or motivated improvement. It does not. Without an answer to who am I trying to become? — the Who Do You Want Your Customer To Become? question turned inward — measurement collapses into surveillance of the self by the self, producing guilt and stasis in roughly equal measure. The Quantified Self stalled for want of three things: a theory of selves as strategic players, a theory of becoming as a directional commitment, and an instrument that produces reflection rather than mere record. AI now plausibly, computationally and economically supplies all three.

This returns the argument to principal-agent theory, but with ‘moral hazard’ relocated. Classical theory assumes a unitary principal and fears a self-interested agent. Computational egonomics inverts the worry: the agent may be perfectly faithful and still serve the wrong principal, because the principal is plural and the agent allies with whichever self holds the prompt. The charter resolves what the classical problem never had to — not the agent's loyalty, but the principal's identity. Designing and discussing ‘agentic AI’ decoupled from this behavioral economic reality is foolish and, yes, counterproductive.

Method. The dieter who buys the ice cream and the dieter who eats it at midnight are different strategic actors, with different information sets, time horizons, and utility functions. ‘Nudge’ Nobelist Richard Thaler's ‘planner-doer’ model proffers the same insight: choice architecture exists precisely because the ‘self that chooses’ and the ‘self that bears the consequences’ are not the same self. Schelling's ‘commitment devices’ — Odysseus lashed to the mast, the hidden cigarettes, the Christmas Club account — were analog instruments of this 1970s paradigm. They 'worked' – sort of - but were static, single-behavior, and left almost no analyzable record of the changes they induced. Schelling’s paradigm was 'correct' but its instrumentation was primitive.

Ongoing ontological shifts cut even deeper: Contemporary digital discourse treats 'the agent' and ‘agents’ as neutral infrastructure that users seek to effectively deploy. Egonomics asserts something edgier: the agent is now a player in the non-cooperative game among a person's selves. In other words, the agent is not merely deployed; it is assigned — by default or by design — to an alliance. That agent booking that impulsive meeting has not served ‘the user;’ it has effectively – affectively? - allied with the 'impulsive self' against the strategic self in a game the principal did not know was running. This paper’s governing – and governance – question: whose ally is the agent, and on what authority? Which ‘self’ runs the show? How do we know? Better yet, how can we measure?

AI’s evolving capabilities radically transform the economics – and revealed preferences – of both computational agents and human agency. A non-cooperative game is empty without its players — and the framework here must pay its own bill rather than borrow the authority of the formalism. Who names the selves? Not the ‘theorist’ by stipulation; that is ‘Quantified Self’ astrology. Nor is there a sovereign self standing outside the portfolio to ratify the rest — that ‘self’ is the oldest fiction in the literature, and looking for it only relocates the problem. Our resolution: stop looking!

Selves are not declared and then approved; they are revealed by enforcement. A ‘self’ exists, for this architecture, exactly insofar as it can ‘bind’ the others: the midnight self is real because it overrides the morning's commitment and the strategic self is real because it builds a ‘commitment ledger’ that overrides the midnight one. The agent roster is therefore not a list the principal certifies; it is the record of which selves have bound which, written by the commitment devices themselves. Operationally, this requires no psychological classification: selves become legible through observable differences among commitments, instructions, overrides, constraints and realized actions. As Amazon founder Jeff Bezos relatedly observed, “Good intentions don’t work, mechanisms do.”

This raises the constitutional question our architecture can’t evade or avoid: if no ‘sovereign self’ stands above the portfolio, who has authority to charter the rest? Computational egonomics does not answer by declaring one self ‘superior;’ it answers procedurally and algorithmically. The charter’s authority derives from the conditions under which it was authored: reflective deliberation, explicit commitment, temporal distance from immediate temptation, and willingness to be evaluated against future outcomes. The architecture privileges procedures, not personalities. It assumes no sovereign self; it makes competing claims of authority visible, revisable, and contestable. The enterprise Ulysses is agentically and algorithmically lashed to his chosen digital mast.

The pivot is conceptual before technical. A person is not a single self with a single utility function; a person is a portfolio of selves, each with characteristic strengths, blind spots, time horizons, and risk appetites. This is a coherent extension of human capital theory under egonomic conditions.

Discussion. This is where current agentic AI deployment paths appear more dangerous. The dominant market incentive is frictionless delegation: the agent handles the task, the recommendation appears ‘automatically,’ the workflow executes invisibly, the preference is inferred before deliberation occurs. This feels ‘efficient’ because it is efficient — and measurably so. But efficiency and agency are not even remotely identical. Paths of least resistance are not highways to maximum advantage.

The critical distinction is not the legacy 'human versus machine' dialectic and trope - and not even 'automation’ versus ‘augmentation' as tidy opposition. The reality: every consequential delegation assigns an alliance. The 'agentic vs agency' question is whether the alliance is assigned by default or by charter and/or constitution. Delegation assigns it by default — to whichever 'self' happened to be present at the prompt. Instrumentation assigns it by charter — to the self the principal has reflectively and/or intentionally chosen to develop. That’s the magic of commitment mechanisms.

For the serious AI-empowered enterprise, this isn't philosophy – it's measurement error with a balance sheet. The productivity case and the agency case are not the same case and – crucially - the second is the one that compounds. A further constraint sharpens this: Agentic AI scales the production of cognitive demands far faster than humans scale their governance of attention. Launching agents is now cheap; reviewing, reconciling, authorizing, and learning from their output is emphatically not. Google’s Addy Osmani calls it ‘the orchestration tax:’ agents parallelize execution, but judgment remains scarce, serial, and bounded. The danger is not Terminator's Skynet; it is developmental drift. And the governance question is not whether a human is ‘in-the loop’ or ‘on the loop’ but whether their judgment is fit for purpose — i.e., which self holds the evaluative lock. Practically and pragmatically speaking, an 'exhausted self' approving twenty agent outputs at a midnight deadline is not the 'strategic self' reviewing three at the dawn of a new day. The CFO of the opening — clearing the queue before the earnings call — is the same person, but not the same governor. Ironically and economically, this is ‘obvious.’

A measurement architecture without a governance architecture is surveillance. The governance problem is, in Leonid Hurwicz's terms, a mechanism-design problem applied to the intra-personal case: given that a person's selves hold private information — each sees what the others suppress — and pursue competing interests, how does one design the agent-mediated interaction so that the dominant strategy across selves yields developmental rather than extractive outcomes?

Posed that way, governance elements are not checklists; they are what the mechanism requires to be incentive-compatible. A charter solves the objective problem: without it, the agent inherits whatever objective the platform optimizes — rarely the principal's — and the alliances are assigned by default. The charter makes purpose computable: it specifies delegated authority, constraints, escalation and reconsideration conditions, override rights, and the purpose against which actions are judged. An Institutional Surprise Rate analog solves the detection problem: a measured capacity to notice when realized behavior departs from chartered direction — in egonomic terms, when unauthorized selves have been playing.

Conclusion. The implications here are not modest: they challenge agent developers to treat the principal's agency as a first-class design target rather than an afterthought. They explicitly recognize that every ‘default’ shipped is effectively a constitutional assignment made on the user's behalf. They also fundamentally challenge enterprises deploying agents at scale, who must accept that productivity and agency are not the same case — and that agency is the one that compounds.

The operative demand is sharper and ruthlessly reframes the central artifact of the whole architecture. Stop writing charters that delegate tasks; craft charters that author ‘selves.’ The charter is no longer a permission slip for the agent — it should be the principal's statement of which ‘self’ they are committing themselves into being and becoming. The three gaps represent the preliminary dashboard of whether the authorship is holding. Perhaps that dashboard will become their GPS.

To be sure, the agents – ever-smarter and more capable - will arrive regardless, thanks to better design or Recursive Self Improvement. Whether they enlarge or erode the humans they ostensibly serve is the singular design decision of the decade. There is no neutral default. The choice is not between human and machine: This choice offers a future in which agents tacitly answer the question of who the human is becoming and a future in which the human, instrumented at last, can actually answer.

Limitations. The strongest and most obvious objection to this deserves discussion: People did not want to be quantified singly and will likely desire quantification of ‘multiple selves’ even less. Relentless deliberation over which self to deploy could become its own agency-erosion, crowding out the unselfconscious actions where much of life's value lives. Agents that simply do things may be more mercy than menace. They may be market winners.

This objection has real and pragmatic force and the reasonable response feels more narrow than triumphant. This architecture is not for every market, domain or domain expertise. For the consequential ones — strategic decisions, professional judgment, long-horizon commitments, the cultivation of capability — the alternative to instrumented selves is not zen-like flow; it is the unmeasured surrender of those decisions to defaults optimized against objectives that are not the principal's own. The effective and affective analogy is sports analytics: baseball without WAR did not feel impoverished to its participants. The measurement was built because the gap between perceived and actual contribution compounded where the stakes were highest. It’s called ‘Moneyball’ for good reason. Most domains need no such instrument. The ones that do get it from nowhere else. That's the disruption.