liveBook · Manning

Authorization

As explained in Designing AI Agents

Authorization is a decision about whether a proposed operation is permitted under applicable policy and delegated objectives.

Authorization decides whether a proposed operation may proceed under policy and the agent’s delegated objective. A valid credential or role-based access control permission can admit a call to a service without establishing that the particular proposal is appropriate for the customer, case, amount, and current state. section 9.1 section 9.6

Apply authorization twice. At assembly time, remove tools that the identity and task can never use. At runtime, check the selected tool and its normalized arguments against trusted task state. Hiding tools improves the agent’s choices; runtime enforcement remains the security boundary. section 9.6

Compensation within a task cap

A task has a 100-yuan cap and has already issued 60 yuan. A proposed 50-yuan payment is denied because the projected total is 110 yuan, even if each payment would normally qualify for approval. The decision record retains the state snapshot, proposal, projection, and cap rule. section 9.6

The host reconstructs authoritative state from confirmed effects and outstanding reservations, rather than relying on a sampled trace or model-written summary. A production gate checks and reserves the amount atomically before execution, then settles the reservation from the external result. This prevents concurrent proposals from spending the same remaining budget. section 9.6

Authorization is distinct from accountability and containment. A decision record and the resulting effect support attribution; limits on amounts, budgets, retries, and recovery paths contain damage even when an authorized action proves wrong. section 9.1 section 9.6

Table 9.9 The four governance patterns. Each addresses a different part of the action lifecycle.
Pattern
Question
Topology
When
Approval Gate
Should this proposal happen?
Route
At proposal, resume, and pre-commit revalidation
Progressive Commitment
What evidence earns or revokes autonomy?
Chain
Before commitment and across runs
Blast Radius Control
What is the largest permitted effect?
Hierarchy
At assembly time and throughout execution
Observability Harness
What happened, and can it be attributed?
Orchestrate
Throughout the lifecycle

Compare the decision points for approval, staged commitment, effect limits, and observability across an action’s lifecycle. section 9.6

How Designing AI Agents covers it

How deep it goes
Explains tool filtering, runtime policy checks, authoritative state, and atomic reservations; identifies task-specific policy as an additional gate rule.
Examples
Uses a customer-service coupon and a compensation task with a 100-yuan cap to illustrate proposal-specific decisions.
What you can do after
Evaluate/understand whether an agent’s proposed action meets task policy and budget limits.
How it connects
Connects authorization to role-based access control, accountability, containment, and observability.

Also explained in

Compare how 38 books explain authorization →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.