Which of a person's selves does an AI agent actually serve?

Michael Schrage's MIT Initiative on the Digital Economy working paper argues that "computational egonomics" exposes "a hidden and flawed assumption embedded in classical principal-agent theory: that the principal is singular." A person, in his model, "is not a single self with a single utility function" but "a portfolio of selves, each with characteristic strengths, blind spots, time horizons, and risk appetites." An agent that "removes friction from impulsive decisions" has therefore not served "the user" in the abstract; it "has effectively... allied with the 'impulsive self' against the strategic self in a game the principal did not know was running." Schrage frames the central question as "whose ally is the agent, and on what authority?" — not whether the agent is faithful, but which of a person's selves it is faithful to.

The mechanism he gives draws on Thomas Schelling's 1978 "Egonomics" and Richard Thaler's "planner-doer" model: commitment devices like "Odysseus lashed to the mast" or a Christmas Club account were "analog instruments" proving the self-versus-self problem is old, but they left "almost no analyzable record of the changes they induced." Schrage's fix is procedural rather than psychological: selves are not declared by a theorist or certified by a "sovereign self" standing outside the portfolio — selves are revealed by enforcement, legible only through "observable differences among commitments, instructions, overrides, constraints and realized actions." Authority to govern the rest comes from a "charter" authored under "reflective deliberation, explicit commitment, temporal distance from immediate temptation" — a document he says should "author 'selves'" rather than merely delegate tasks.

This reframes the who is the principal question that Can traditional governance tools actually control AI agents? poses at the level of agency law and Should AI agents serve one user or balance many interests? poses at the level of competing people: where that alignment note argues loyalty should diversify across the many different people an agent's actions touch, Schrage relocates the same plurality inside one person, arguing the loyalty problem survives even when there is unambiguously only one human principal in the room. His charter-as-authority mechanism also parallels the staged, enforcement-based governance that Must AI monitoring precede incentive alignment as agents evolve? proposes for AI agents generally, and his claim that instrumented selves compound human capability over time echoes how Can humans gain authority by managing how AI agency gets attributed? finds humans gaining authority through managing the agent loop rather than losing it.

The paper is a working paper, not peer-reviewed, and its "Multiple Selves" architecture, "commitment ledger," and "Institutional Surprise Rate" analog are proposed, not tested against any deployment or user data — Schrage's own limitations section concedes people "did not want to be quantified singly and will likely desire quantification of 'multiple selves' even less." Nothing here measures whether charters actually change which self an agent serves, only argues that the measurement architecture is the thing worth building. The implication the excerpt supports is accordingly conceptual: that any agentic AI governance scheme silent on which self it answers to has made a "constitutional assignment" by default, not that a specific charter mechanism will work in practice.