X · Mēden agan · Nothing in excess
Authority stays human.
Automation should never outrun accountability.
Explicit permissions and approval on every consequential action.
In practice
What it means for your system.
Agents prepare, draft, and act within set limits. Decisions with consequences stay with a named person.
- Each agent has explicit permissions, and nothing beyond them.
- Consequential actions wait for a named approver.
- Every approval is recorded, so accountability stays visible.
The word for it
Phronēsis
φρόνησις
Phronēsis is Greek for “practical wisdom, judgment about what to do.” People keep the judgment calls, and agents prepare the work.
Why it matters
The agent carries the load. A person holds the seal.
An agent can act faster than anyone can notice a mistake. The liability for that action still belongs to your business, and to the people who answer for it.
Approval on every step wastes the time agents are meant to save. Approval on no step leaves consequential actions without an owner.
So authority is sized to consequence. Routine work moves freely within set limits, and the actions that commit money, customers, or records wait for a named person.
Read it accurately
Scope of the rule.
- Each approval belongs to a named person.
- Each approval belongs to a named person with the authority to make it. The record shows who approved each action, and when.
- Limits are enforced outside the agent.
- Permissions are enforced by the system, outside the agent. An action beyond its limits is held, whatever the agent concludes.
- People hold the accountability.
- Logs and controls support accountability, and people hold it. A named person answers for each consequential action an agent takes.
How it is measured
Every rule is something you can check.
Explicit permissions, approval boundaries, and controls on every critical action.
In an engagement
Where the rule is applied.
The rule is checked at each stage of the work, from the first design to the system in operation.
In the design
Every action each agent could take is listed and sized by consequence. A System Blueprint names the approver, and a backup, for each consequential action.
In the build
Custom Agentic Systems enforces permissions in the system and holds consequential actions for the named approver. Each approval is recorded with its evidence.
At acceptance
Tests attempt actions beyond each agent’s permissions and confirm they are held. Your owner signs off on the approval boundaries before release.
In operation
The approval record is reviewed for patterns over time. Your owner decides whether any boundary should move, and nothing widens by default.
Check your own system
Five questions to ask this week.
Each one has a yes or no answer. A no marks where to start.
- Can you list every action each of your agents is permitted to take?
- Are actions that commit money, customers, or records held for approval?
- Is a named person, with a backup, responsible for each approval?
- Can you show who approved a given action, and when?
- Are an agent’s permissions reviewed whenever the agent changes?
Delivered through
The services that apply this rule.
Mēden agan
Nothing in excess.
No more model, spend, or autonomy than the work requires.
Apply the rule
Check it against your own system.
Questions
Doesn’t approval slow the work down?
Approval is placed only on consequential actions. The agent prepares everything else, so the approver decides with the evidence already assembled.
Can an agent’s permissions grow over time?
Yes, when the record supports it. Your owner reviews the evidence and approves each expansion, and nothing widens by default.
What about agents we already run?
Agent Assurance adds permissions, approval boundaries, and tracing to agents already in production. It keeps those controls enforced as the agents change.
How do we check this rule on our own system?
Pick a critical action and trace its controls. It should have explicit permissions, a clear approval boundary, and a control that holds it for a person.
What happens when the approver is unavailable?
The action goes to the named backup. If neither is available, it waits, and the agent continues with work that needs no approval.
Which actions count as consequential?
Actions that commit money, customers, or records, or that are hard to reverse. Your owner confirms the list during design, and it is reviewed as the system changes.