← Patterns

Capability-Autonomy-Risk Triangle

Core Concept

For AI agents, you can optimize for two of three: capability, autonomy, risk. You cannot have all three.

High capability + High autonomyHigh risk
High capability + Low riskLow autonomy (human checkpoints, narrow scope, mandatory logging)
High autonomy + Low riskLow capability (the agent can’t do much that matters)

This is the “good / fast / cheap, pick two” of the AI era. It’s the planning framework for deploying agents inside organizations.


The Logic

A capable agent acting freely will eventually do something costly. A minimal agent acting freely is fine — but it doesn’t move the needle. The only way to combine capability with safety is to trade autonomy for human/system checkpoints.

“A highly autonomous agent incurs more risk for the organization the more capable it is. A minimal agent is able to act freely with little risk.” — Robbie McKinstry, DevGuide conference takeaway (May 2026). See source.


How to Apply

When deploying any AI agent (Claude Code, Codex, in-house agents, vendor agents), ask:

  1. What’s the blast radius if it does the wrong thing? (Estimates risk.)
  2. How capable does it need to be to deliver value here? (Sets the capability target.)
  3. How much autonomy can I give it given (1) and (2)? (Forces explicit autonomy choice rather than default-high.)

Then choose the structural lever:

GoalLever
Reduce risk without losing capabilityAdd checkpoints, narrow scope, mandatory logging, dry-run modes
Increase autonomy without raising riskMake the action reversible, sandbox it, or pre-commit to constraints (Autonomy-Through-Constraints)
Raise capability without raising riskSplit the work — high-capability plan, low-autonomy execute (Build-AI-Run-Deterministic)

The triangle explains why “AI doesn’t replace lawyers” reads wrong. The right framing is: capability is already there; the question is how much autonomy a firm can stomach. Low-autonomy + high-capability AI (great research assistant) is shipping today. High-autonomy + high-capability AI (acts on the client’s behalf) is gated by risk, not by model quality.


Where I’ve Seen It



Cross-References