Principal agent

Principal–Agent Problem

Overview

One party (the principal) delegates to another (the agent) whose interests differ and whose actions can't be fully observed — producing agency cost: monitoring spend + agent bonding spend + residual loss. Formalized by Jensen & Meckling (1976). Structure produces the behavior, not character — so the fix is structural.

Composes with signaling-games, repeated-games-reputation, prisoners-dilemma, and okr-goal-setting.

When to Use

• Board reviewing executive compensation; outsourcing or contractor decisions
• Employees/executives behaving in ways that puzzle leadership
• New joint venture, LP-GP fund, or platform marketplace being structured
• Someone says "agency cost," "moral hazard," "skin in the game," "fiduciary duty"
• Deploying an autonomous AI agent, sizing AI capex/adoption, or facing AI-native competition where you delegate to a system whose objective and actions you can't fully observe (alignment / guardrails / human-in-the-loop)

Not when: fully aligned interests + fully observable behavior; contract design cost exceeds the agency cost it would prevent.

Coaching Novices (Adaptive Front Door)

• Engine mode: concrete case → run The Process directly.
• Coach mode: unfamiliar or no concrete case → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. One-liner: when one party delegates to another whose interests differ and actions are unobservable, the agent will systematically act in ways the principal didn't want — cure is structure, not character.
2. Check fit: fully aligned + fully observable → no agency problem.
3. Elicit their specific relationship — who is principal, who is agent, what does each really want?

[WAIT — do not advance until user responds]

4. Probe: what can the principal not observe? which misalignment dominates (effort / risk / time horizon / info asymmetry / multitasking)?

[WAIT — do not advance until user responds]

5. Close: name the specific misalignment and one structural lever (incentive, observability, or selection).

[WAIT — do not advance until user responds]

The Process

Step 1 — Identify structure Principal / Agent / What principal wants / What agent would do absent intervention / What principal cannot observe.

Step 2 — Diagnose misalignment 1-3 dominant types: effort · risk · time horizon · info asymmetry · adverse selection · moral hazard · multitasking · hidden self-dealing.

Step 3 — Estimate agency cost Monitoring cost + bonding cost + residual loss = total. Order-of-magnitude is enough.

Step 4 — Design alignment mechanisms (a) Incentives: equity, performance bonuses, carried interest, profit-sharing, skin in the game. (b) Observability: audits, reporting, independent verification, public reputation systems. (c) Selection: reference checks, work samples, trial periods, self-selection through contract design.

Step 5 — Trade off — optimum minimizes the sum of all three costs, not any single one.

Step 6 — Accept residual cost — quantify it, decide if acceptable, build into forecasts.

Output: Principal-Agent Analysis

# Principal-Agent Analysis: <relationship>
- Principal: / Agent: / Delegated task: / What principal cannot observe:
- Primary misalignment(s): / Estimated agency cost:
- Incentive mechanism: / Observability mechanism: / Selection mechanism:
- Residual cost: <amount> — Acceptable?: <yes/no>
- What changes about how we structure this:

→ Method in Action: Jensen-Meckling 1976 and the Enron Collapse, 2001

→ 2026 lens: Delegating to an Autonomous AI Agent (2024–2026)

Pack: Common Patterns (see also full table in examples)

Shareholders ↔ CEO (stock gaming) · Investors ↔ Fund manager (AUM vs returns) · Company ↔ Sales reps (discount-to-close) · Patient ↔ Doctor (procedure-volume billing) · Client ↔ Attorney (hourly billing) · Platform ↔ Users (rent extraction).

Applying It Well

• Use multiple mechanisms: incentives + observability + selection. Mono-mechanism designs are brittle.
• Agency problems compose multiplicatively across levels — analyze each boundary separately.
• Most dangerous relationships are the unrecognized ones (framed "fiduciary" when the structure says otherwise).

→ Primary sources: references/sources.md

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Fake move	Reality
[D] "We just need to hire good people"	Structure produces behavior; good people in bad structures behave badly.
[D] "Our agent has skin in the game" (small stake)	Size matters — 1% equity barely shifts behavior.
[D] "We trust them"	Trust without structural alignment is the bonding mechanism the structure exploits.
[D] "We have an oversight committee"	Captured or info-starved boards don't constrain agents. Enron's board met regularly.
[D] "Long-term incentives align them"	Most "long-term" plans vest at 3-4 years — short relative to many decision horizons.
[D] "Performance metrics solve agency"	Agent optimizes the metric; principal's real interest decays (Goodhart's law).
[D] "This is a fiduciary relationship"	Legal duty adds recourse after the fact; structural alignment still needs designing.
→ Add [O] entries here after each real use — paste the actual failure pattern	What went wrong and why

Red Flags

• P-A relationship exists but not named or analyzed; compensation purely fixed for an outcome-sensitive role
• Observability poor — agent's behavior cannot be measured
• "We trust them" used as substitute for structural alignment
• Board or audit body captured by the agents it supposedly oversees

Verification

• Principal and agent explicitly named
• What principal cannot observe is stated
• Specific misalignment(s) diagnosed
• Agency cost estimated (order-of-magnitude is enough)
• At least one mechanism in each of incentive, observability, selection
• Trade-offs across mechanisms acknowledged
• Residual cost accepted explicitly, not assumed away
• Exit mechanism for the principal preserved

---

Part of deciqAI Knowledge Skills — 233 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/s/principal-agent · Built by deciqAI · github.com/deciqAI · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/principal-agent.json