Chapter 3.2 — Agency, limited autonomy and blast radius¶
🎯 Objective¶
Establish that autonomy is not an emergent property: it is a function of the action space, the available tools, the granted permissions and the applied policies. Whoever designs the agent designs the ceiling of autonomy, and therefore the ceiling of risk.
🧠 Concepts¶
- Agency. The operational capacity to pursue a goal via intermediate decisions and permitted actions.
- Limited autonomy. Agency constrained by scope, available tools, IAM permissions, cost budgets, a maximum number of steps and human approvals at critical points.
- Blast radius. How much a wrong action by the agent can damage the system/company. Includes reversible effect, permanent effect, number of tenants affected, financial value exposed.
🛡️ Principles¶
- Fewer permissions > more permissions. Increasing privilege is a conscious decision, with review; reducing is the default.
- Human approval for irreversible actions. Objective criteria (value, action class, affected resource) declared before go-live.
- Budget + circuit breaker to limit cost and number of steps.
- Observability requires tracing of each decision, not just the final result.
- Defense in depth across identity, policy, schema and tool registry — no isolated layer.
🚨 Signs of excessive agency¶
- An agent with access to tools it "does not need yet", but "might come to need".
- An agent sharing a service credential with a human or with another agent.
- An agent without a clear
max_steps. - An agent without a per-tenant budget.
- An agent whose only defense against misuse is the prompt.
📚 References¶
- Anthropic — Measuring AI agent autonomy in practice: https://www.anthropic.com/news/measuring-agent-autonomy
- OWASP LLM Top 10 — LLM08 Excessive Agency: https://owasp.org/www-project-top-10-for-large-language-model-applications/