Diop Daily #114 — September 2026

The Agent Needs a Theory of Its Principal

An agent that works for someone needs more than a task description. It needs a working theory of the person or institution whose purposes it serves: what that principal is trying to protect, which trade-offs it accepts, how it speaks, when it abstains, and which decisions require an explicit handoff. This theory is the difference between a machine that produces plausible output and one that can carry delegated work without quietly changing its meaning.

OpenAI’s News RSS reported on September 14 that Fyxer uses OpenAI models, fine-tuning, memory, and real user feedback to organize inboxes and draft emails in each user’s voice. This is a vendor-reported description of one product and does not establish general performance. It does identify a useful direction: the assistant is being shaped by an ongoing relationship with a principal, not only by the contents of a single prompt.

A faithful agent carries the shape of the person it serves without pretending that resemblance is authority.

A mandate has a living interior

Delegation is often represented as a sentence: “handle my inbox,” “prepare the report,” “respond to this customer,” or “run the analysis.” The sentence names a task while leaving the principal’s interior mostly unstated. Which messages deserve immediate attention? Which relationships should be protected? Which claims require evidence before they leave the institution? Which tone is appropriate for a regulator, a colleague, a customer, or a family member? What should the agent leave untouched even when it has the technical ability to act?

People answer these questions through accumulated context. They observe one another’s corrections, notice repeated preferences, learn what counts as a dangerous shortcut, and infer when silence is better than a polished response. An agent must make that learning more explicit. If its model of the principal remains hidden inside weights, transient context, or an opaque vendor profile, the relationship cannot be inspected when a consequential action is disputed.

The principal model should therefore be treated as a governed object. It can contain stable commitments, current priorities, audience-specific conventions, prohibited shortcuts, examples of accepted work, examples of rejected work, escalation rules, and the evidence behind each important change. Some elements describe identity. Others describe policy. Others are temporary instructions whose validity expires with a project or decision. Keeping these categories distinct prevents a passing preference from becoming a permanent command.

Calibration is the work between prompts

A principal does not teach an agent by issuing instructions alone. The principal teaches through corrections. A revised paragraph says that the emphasis was wrong. A rejected recommendation says that a risk was underweighted. A changed recipient says that the social boundary was misunderstood. A request to preserve a phrase may reveal a legal, cultural, or historical commitment that the machine could not infer from style.

Those corrections are evidence about the relationship, yet they require interpretation. A single edit may be local. Ten similar edits may reveal a durable preference. A correction made during a crisis may not belong in ordinary operations. A preference expressed by one staff member may not represent the institution. Calibration needs attribution, scope, confidence, and expiry so that learning remains accountable.

A useful calibration record can carry:

  • Principal identity: the person, team, institution, or role whose purposes the agent is serving, with the authority that establishes the relationship.
  • Operating context: the workflow, audience, jurisdiction, language, time period, and external systems to which the preference applies.
  • Observed evidence: the instruction, correction, accepted output, rejected output, or decision that supports the inferred preference.
  • Strength and scope: whether the rule is a hard boundary, a strong preference, a local convention, or a hypothesis awaiting more evidence.
  • Abstention condition: the situation in which the agent should pause because the principal model is uncertain, conflicting, stale, or outside its authority.
  • Review path: who may confirm, amend, revoke, or export the calibration record and how dependent work is located when it changes.

This record allows an institution to ask a precise question after an error: which assumption about the principal produced this action, what evidence supported it, and who can correct it? The answer is more useful than a vague appeal to personalization. It turns adaptation into something a reviewer can examine and a successor system can inherit.

Voice is a boundary, not a costume

Writing in someone’s voice is often treated as a problem of imitation. The deeper issue is representation. A message carries the sender’s priorities, relationship to the recipient, tolerance for uncertainty, and willingness to make a commitment. A sentence can sound accurate while committing the principal to a position that was never authorized.

Voice calibration must therefore include abstention. The agent needs examples of the principal’s language, but also examples of subjects that require a pause, subjects that demand a source, and subjects that belong to another role. A public statement may require institutional approval. A research claim may require a source trail. A customer response may require a price or service check. A personal message may require the person’s direct attention. Style cannot settle these questions by itself.

This is where representation differs from impersonation. Impersonation hides the gap between machine output and human authorization. Representation keeps the gap visible and gives the agent a disciplined way to stop before crossing it. The aim is faithful assistance within a declared relationship, with a clear boundary around decisions the machine may prepare but not finalize.

Feedback needs a chain of custody

Google’s public description of ADK Go 2.0 presents graph-based workflows, human-in-the-loop orchestration, dynamic routing, retries, and built-in resilience. These primitives make it possible to route a task and request a human decision. They do not determine whether the correction should alter the principal model, which principal it belongs to, or whether a local adjustment has wider institutional scope.

That distinction becomes important when several agents work together. One agent may classify messages, another may draft a response, and a third may check policy or evidence. If the drafting agent learns from a correction without recording the task context, the correction may leak into unrelated work. If the checking agent rejects an output for a legal reason and the system stores the rejection as a style preference, later drafts will be calibrated against the wrong lesson.

Feedback should move through an accountable chain:

  1. Capture the original instruction, machine output, human correction, workflow context, and affected principal before interpreting the lesson.
  2. Classify the correction as content, style, policy, authority, risk, language, or temporary circumstance.
  3. Test whether the correction applies to one artifact, one workflow, one role, or the principal’s broader operating model.
  4. Store the resulting rule with its evidence, confidence, scope, expiry, and reviewer rather than replacing the prior record silently.
  5. Replay representative cases before allowing the new calibration to govern external communication or other consequential actions.

The chain protects both sides of the relationship. It prevents the agent from forgetting useful learning, and it prevents the principal from becoming trapped by an unexamined inference made from one moment. A system that remembers every correction without classifying it accumulates noise. A system that remembers none of them forces the institution to teach the same lesson repeatedly.

Governance begins with attribution

NIST’s AI Risk Management Framework gives this problem a practical vocabulary through Govern, Map, Measure, and Manage. Govern identifies who owns the principal model and who may approve changes. Map identifies the people, records, tools, audiences, and obligations affected when the agent acts on an inferred preference. Measure records calibration drift, correction frequency, abstention quality, and the cost of reviewing a misrepresented action. Manage changes the model, the rule, the route, or the authority boundary when evidence shows that the current relationship is unsafe or unfaithful.

Several measurements are especially revealing:

  • the proportion of drafts accepted without substantive correction, separated from cosmetic edits;
  • the rate at which the agent correctly abstains when principal intent is ambiguous;
  • the time required to trace a disputed output back to the calibration evidence that shaped it;
  • the number of corrections that were wrongly generalized across roles, languages, audiences, or workflows;
  • the percentage of the principal model that can be exported, reviewed, and restored after a provider or model change.

These measures make personalization legible as an operating capability. A high acceptance rate can hide a principal who has stopped reviewing because review is too expensive. A low correction count can mean good calibration, or it can mean that the agent no longer exposes enough of its assumptions for correction. The quality of representation is measured by the relationship between output, authority, evidence, and repair.

For African institutions, the principal is often collective

Many imported assistant patterns assume a single user with a stable set of preferences. African institutions frequently operate through layered authority: a director, a working team, a community mandate, a public obligation, and a set of language or cultural commitments that cannot be reduced to one person’s style. A cooperative, archive, university, clinic, newsroom, or civic office may need an agent to serve several principals at once, each with a different right to instruct, review, or veto.

The principal model must represent this plurality. A local-language correction may carry an institutional meaning that a translation system would flatten. A community archive may distinguish the authority to read from the authority to publish. A university laboratory may let a researcher explore a dataset while reserving decisions about release to a committee. A public office may permit a service agent to prepare a response while keeping the final obligation with a named official.

Intellectual sovereignty includes the ability to define these relationships in the institution’s own terms. Local control of the calibration record matters because it preserves the history of how a community, role, or organization taught its machine representative to act. External models can contribute capability while the principal relationship remains governed locally, in the languages and categories that make the institution intelligible to itself.

Where the investable surface is widening

If agents are moving from one-shot answers toward recurring delegated work, capital should examine the infrastructure that makes principal calibration explicit and transferable:

  • Principal-model registries: systems that store identities, roles, commitments, preferences, abstentions, evidence, scope, and expiry as reviewable records.
  • Calibration workbenches: tools that turn accepted and rejected outputs into classified lessons, test cases, and controlled updates instead of opaque personalization.
  • Representation and abstention evaluators: suites that measure whether an agent preserves intent, recognizes ambiguity, respects audience boundaries, and stops before unauthorized commitment.
  • Feedback lineage systems: infrastructure linking a correction to the output, workflow, reviewer, rule change, and dependent artifacts that may need replay.
  • Portable relationship layers: exportable principal models that survive a change of model, runtime, provider, language, or institutional operator.

The underwriting question is concrete: does the product reduce the cost of teaching an agent to represent a principal faithfully while making errors easier to detect and repair? Evidence should include correction-adjusted acceptance rate, false-generalization rate, abstention precision, calibration replay time, reviewer minutes per update, cross-language fidelity, and restoration success after a provider change.

This commercial object is distinct from permissions, storage, workflow routing, provenance, and legal signature. Permission determines what an agent may do with a record. Storage preserves the state of the work. A workflow graph routes activity. Provenance records how an artifact changed. A signature or charter establishes authority for a decision. Principal calibration models how a legitimate agent should interpret the principal’s purposes, style, limits, and corrections before it prepares or recommends an action.

Build the relationship before widening the delegation

An institution can begin with a low-consequence workflow that still contains real judgment:

  1. Name the principal, the agent’s role, the audiences involved, and the decisions that remain human.
  2. Collect accepted outputs, rejected outputs, corrections, abstentions, and representative edge cases without collapsing them into one generic preference profile.
  3. Classify each lesson by scope, authority, language, evidence, confidence, and expiry.
  4. Run a replay set that includes ordinary cases, ambiguous cases, adversarial cases, and cases where the correct action is to ask.
  5. Export the calibration record and let an independent reviewer determine whether another agent could understand the relationship without access to the original model provider.

Only then should the agent receive a wider communication or execution interval. A principal deserves a machine that can learn its working logic while keeping the source of that learning visible. The agent’s resemblance may open the relationship, but its accountable theory of the principal is what allows the relationship to endure.

Sources