Field Note · Applied AI and system ownership
Does This Work Need an AI Agent or a Workflow?
Agent is becoming a label people attach to almost any AI process. I would start with a less exciting question: what is the simplest reliable operating shape that can produce the outcome and show us when it is wrong?
A thinking frame by Andrew Moss
The questions I get
Usually some version of these:
- Should we build an agent for this?
- How much autonomy should the system have?
- When is a conventional workflow the better answer?
What a lot of people seem to think
The more capable the model, the more autonomy the process should receive.
How I look at it
Capability is not authority. I use five tests before adding autonomy: recurrence, verified context, consequence, human control, and measurable outcome. A fixed workflow is often better when the path is known. Agent behavior earns a place when the path must adapt and the boundaries, approvals, and evidence remain clear.
Why the decision matters
The cost is rarely confined to the line item.
If the sequence is wrong
An impressive demo can become an unobservable process that takes consequential action from incomplete context.
If the sequence is right
The system uses only as much flexibility as the work requires, with clear inputs, approvals, logs, exceptions, and outcome measures.
How reversible is it?
Depends on the action. A draft is easy to reverse; a message, payment, permission change, or external commitment may not be.
The short answer
Choose an autonomy rung, not an agent label.
Start with assist, then recommend, then prepare action, then execute only bounded reversible steps with approval and evidence. Move up only when recurrence, verified context, consequence controls, human authority, and outcome measurement support the next rung.
The five-part placement testRecurrence · verified context · consequence · human control · measurable outcome
If the work does not recur, context cannot be verified, consequences are hard to contain, the human decision is unclear, or outcomes cannot be measured, more autonomy is usually premature.
Move fromNovelty-first architecture→Move towardThe simplest reliable operating shape
The order I would use
Take the right steps in the right order.
- 01
Name the endpoint
Define the final artifact, recommendation, or action and the standard it must meet.
- 02
Map the stable path
Separate steps that are deterministic from steps that require interpretation or adaptation.
- 03
Classify consequences
Mark what is reversible, externally visible, permission-changing, financial, confidential, or time-critical.
- 04
Place human control
Assign who reviews, approves, transmits, overrides, and owns the outcome.
- 05
Instrument the result
Log sources, decisions, exceptions, approvals, and outcome evidence so the system can be tested and improved.
Questions worth answering
Before the next irreversible move:
- Does the path truly need to adapt, or is the input merely variable?
- What can the system do without creating an external commitment?
- Who has authority to approve the next rung?
- How will we know the system helped rather than merely acted?
What not to do
Do not confuse fluency with reliability.
Do not make an agent because the word sounds advanced. Do not hide a fixed process inside an unpredictable loop. Do not let the system transmit, pay, grant access, or make a consequential choice without explicit authority.
Keep the perspective
The mature design is often quieter than the demo.
The goal is not maximum autonomy. It is dependable leverage that preserves expert judgment and makes failure visible early.
Independent sources
Useful primary material
These sources support the public frame. They do not replace the private facts or the accountable professional.
Common follow-up questions
Is every multi-step AI process an agent?
No. A workflow can contain model-assisted steps while the path and controls remain fixed.
When should autonomy increase?
After the lower rung has reliable inputs, known failure modes, meaningful evaluation, clear authority, and a reason the adaptive behavior improves the outcome.