Decision Notes

Agents don't pose a technical problem. They pose a signature problem.

When an agent decides, who assumes responsibility? The question isn't technical, it's political.

agents responsibility decision

Real situation observed

Demos of autonomous agents show systems that make decisions, execute actions, modify states. The dominant illusion is that this autonomy solves a technical problem: speed, scale, cost reduction.

Dominant illusion

Technical autonomy is presented as an end in itself. We talk about “delegation to the agent” as if it were delegation to a colleague.

Real mechanism

An agent is not an employee. It’s a program that executes instructions. When it “decides”, it applies rules defined by humans. But these rules don’t cover all cases. Unforeseen cases create unsigned actions.

Cause → Effect:

  • An agent acts without validation → An unforeseen consequence occurs → No one signed this action → Responsibility is diluted

What breaks if we ignore the signal

  1. Diluted responsibility: When an agent makes a decision that causes damage, who is responsible? The developer? The product owner? The CTO? The company?

  2. Legal risk: In case of litigation, the absence of human signature creates a gray area. Insurance, contracts, regulations are not adapted to unsigned decisions.

  3. Loss of organizational control: Decisions made by unsigned agents escape hierarchical control. Impossible to trace who authorized what.

Who bears final responsibility

The executive who authorizes the deployment of an agent without signature. Not the developer. Not the agent. The executive.

Decision to make

Every agent action that modifies an external state (database, API, email, transaction) must be signed by an identified human before execution.

Assumed cost:

  • Added latency (human validation)
  • Operational cost (validation time)
  • Technical complexity (signature system)

Benefit:

  • Total traceability
  • Clear responsibility
  • Regulatory compliance
  • Risk reduction

Implementation

  1. Explicit signature: Before any critical action, the agent submits a signature request with context.
  2. Traceability: Each signed action is recorded with: who, when, why, context.
  3. Escalation: If the signature doesn’t arrive within a defined timeframe, the action is canceled or escalated.

This is not a technical constraint. It’s an assumed organizational constraint.

DECISION LEDGER

1. Décision explicite

Every agent action must be signed by an identified human before execution.

2. Ce que cette décision optimise

Clarity of responsibility chain, traceability of decisions, reduction of legal risks.

3. Ce que cette décision sacrifie

Execution speed, apparent agent autonomy, scalability without friction.

4. Niveau de réversibilité

Semi-réversible

5. Qui signe cette décision

CTO in coordination with Legal

Cet article devient faux ou dangereux si :

  • If regulation requires total agent autonomy
  • If the cost of human validation exceeds the value created