Authority
What may the agent do?
Define its authority action by action: observe, recommend, prepare, decide, and execute.
ENTERPRISE AI AGENT OPERATING MODEL
A.G.E.N.T.S.
A practical operating model for moving enterprise AI agents from experimentation into governed production.
Created by Jim Markunas
Enterprise agents create a different product problem from chatbots and conventional automation.
Once software can evaluate changing conditions, choose among permitted actions, interact with business systems, and create real-world consequences, model capability is only part of the design problem.
A.G.E.N.T.S. provides six operating questions for defining that responsibility before an agent is trusted with real work.
“A chatbot can give you an answer. An agent can create a consequence.”
Readiness Console
A.G.E.N.T.S. Production Readiness Check
Interactive assessment
Assess whether your AI agent is production-ready.
Start with the enterprise foundation, define the business value, then assess each A.G.E.N.T.S. control.
Assessment format
9 steps · guided review
The A.G.E.N.T.S. Operating Model
Guardrails
What must constrain it?
Define the policies, approvals, limits, prohibited actions, and hard stops that bound its authority.
Evidence
How will we know what it did?
Preserve enough operational evidence to reconstruct consequential decisions and actions.
Network & Integrations
What systems may it touch?
Define the systems, APIs, identities, data, permissions, and capabilities the agent may use.
Transfer & Escalation
When does a person take over?
Define when the agent must stop, where the work goes, what context transfers, and whether the agent may resume.
Success & Accountability
Did it create value — and who owns the result?
Measure the business outcome and identify the human owner accountable for that outcome.
The problem A.G.E.N.T.S. addresses
Most discussions about AI agents begin with model capability: reasoning, memory, tools, orchestration, or autonomy.
A.G.E.N.T.S. starts somewhere else:
What happens when an organization delegates real authority to software?
Once an agent can change an order, issue a refund, update a system of record, contact a customer, initiate a workflow, or make another consequential decision, the organization also has to define the agent’s authority, boundaries, evidence requirements, system access, escalation model, success measures, and accountable human owner.
This is why A.G.E.N.T.S. treats an enterprise agent as a governed digital product rather than simply a model connected to tools.
From Opportunity to Production
A.G.E.N.T.S. sits inside a simple four-stage implementation path.
FINDDEFINEGOVERNPROVE
FIND
Identify the business opportunity, current process, relevant systems, friction, and potential value.
DEFINE
Define the agent’s objective, information needs, decisions, actions, systems, and expected outcome.
GOVERN
Apply the six A.G.E.N.T.S. controls before consequential authority is delegated.
PROVE
Measure whether the agent produced the business result it was designed to improve.
Authority is a ladder, not a toggle
OBSERVERECOMMENDPREPAREDECIDEEXECUTE
An enterprise agent does not need one universal level of autonomy. Authority should be assigned action by action according to consequence, reversibility, operating rules, and the value of human review.
“Autonomy should be granted action by action, not agent by agent.”
Worked Example: Customer Order Exception Agent
Consider an agent responsible for investigating delayed or failed customer orders across customer, order-management, inventory, and payment systems.
It may inspect the current state, determine which approved resolution applies, execute bounded low-risk actions, and escalate higher-risk or conflicting cases.
Refund ≤ $50
May execute
Refund > $50
Prepare for human approval
Payment already reversed
Never refund again
Conflicting payment state
Stop and escalate
The purpose of A.G.E.N.T.S. is to turn decisions like these into an explicit operating model before production deployment.
Evidence should reconstruct the consequence
For consequential actions, an organization should be able to reconstruct what information the agent used, what state existed, which constraints applied, what decision was made, what action occurred, whether a human intervened, and what happened afterward.
SOURCESTATECONSTRAINTDECISIONACTIONOVERRIDEOUTCOME
“Audit the system’s behavior, not just its explanation of its behavior.”
Success starts with the business case
Agent activity is not the outcome.
The implementation should return to the value case that justified the agent in the first place.
Increase ROI / Revenue
Decrease Cost
Operational Efficiency / Streamlining
SUCCESS METRICBEFOREAFTERIMPROVEMENT
If the metric the agent was created to improve does not materially improve, AI activity alone is not evidence of business value.
A.G.E.N.T.S. Resources
About the Framework
A.G.E.N.T.S. was created by Jim Markunas as a practical operating model for product teams designing enterprise AI agents that can make decisions, use tools, interact with business systems, and create operational consequences.
If any A.G.E.N.T.S. control is unclear, the agent is not ready.
A — What may it do?
G — What must constrain it?
E — How will we know what it did?
N — What systems may it touch?
T — When does a person take over?
S — Did it create value, and who owns the result?