Decision guide
AI agents vs deterministic flows
«Agent» now gets used for any system with a model inside, and there is a real difference underneath. This guide reduces it to one question, gives you a matrix to apply it to your own processes, and explains why the answer is almost always a flow.
The short answer
Use a deterministic flow when you can list the steps in advance. Use an agent when the number of steps depends on what turns up along the way. Most production systems should stay flows, with a bounded stretch of autonomy where it is needed.
Updated on
The difference: who decides the path
It is not how much AI the system contains. Anthropic's Building Effective Agents draws the line at a single question. In a workflow, the model and tools are orchestrated through paths written in code beforehand. In an agent, the model directs its own process and chooses which tool to use.
Decision matrix
| If this is true | Deterministic flow | Agent |
|---|---|---|
| You can list the steps in advance | Yes | No: you are paying for flexibility you do not need |
| A wrong answer reaches a customer unreviewed | Yes, with verified output | Only behind a verification gate |
| The number of steps depends on what turns up | It will fight you | Yes: this is the case the pattern exists for |
| You need to explain a specific past run to someone | Yes | Harder: the path was chosen, not written |
| The cost per run has to be predictable | Yes | No: the model decides how much work to do |
Start with simple prompts, optimize them with comprehensive evaluation, and add multi-step agentic systems only when simpler solutions fall short.
Anthropic, Building Effective Agents
Where a canvas runs out
The usual claim is that visual builders break past some number of nodes. That threshold is not documented anywhere. What you can check is the mechanism.
A canvas holds the happy path very well. It holds badly whatever is not a path: an invariant («this figure always comes from the quote»), a contract («this field has a maximum length») or a policy («never answer a price without a source»). None of those are steps, so they have no place on a diagram and end up as more branches.
The real tell is not the node count. It is the day the team starts adding sub-workflows and code nodes: that is the canvas admitting it stopped being the representation of the system.
Bounded autonomy: contracts, not trust
If you give a model room, the useful question is not how much freedom it has but what it cannot do regardless. Three limits do most of the work.
Validate input before anything decides
What reaches the decision code already has a known shape, checked at the boundary.
Give every capability a contract
What it takes, what it returns, and for which customer. A capability that checks who is asking cannot be talked into answering for someone else.
Verify output before it leaves
Every figure in the final message must come from an authority. One that does not is discarded and a deterministic template answers.
The shape that works: a flow that delegates
The two are not exclusive. A deterministic spine handles the turn, the money and the contracts; a bounded agent takes the sub-task whose length nobody can predict, such as research or reconciling a messy document, and hands back a typed result that the spine verifies like any other input.
The full reasoning is in Workflows vs agents, and the product stance in the manifesto.
The map of the topic
- Agent protocolTyped tools, scoped permissions and a person on the write path.
- Token costLinear flows against agent orchestration, in numbers.
- Schema driftWhy structured extraction fails silently.
- Bentho vs n8nWhen to stay on n8n and when to move a flow with a model.
- Bentho vs MakeWhat changes when the answer is not deterministic.
- Architecture auditA diagnosis of where what you have breaks first.
Frequently asked questions
What is a deterministic flow?
A process with its steps written in advance: the same input follows the same path. It can have a model inside, as long as the code chooses the route.
When do you need an AI agent?
When you cannot list the steps because they depend on what turns up along the way, such as researching a case or reconciling documents that disagree.
Is a flow with a model inside an agent?
No. If the model fills in values and the code decides the route, it is still a flow. It is an agent when the model chooses what to do next.
Can flows and agents be combined?
Yes, and that is usually best: a deterministic spine for money and contracts, and a bounded agent for the unpredictable sub-task, with its result verified on return.
How many nodes can a canvas take before it breaks?
There is no documented number. The real limit shows up when rules that are not steps, such as invariants and policies, turn into more branches.
Apply the matrix to your process
If you have a flow with a model inside and cannot tell which side it falls on, tell us how it is built and we will look at it with you.
Where the data comes from
What this page says about other products comes from their public documentation. It may have changed: if you are about to decide on a figure, check your own.
- Anthropic, Building Effective Agents (checked on September 25, 2026)
It separates workflows and agents by who decides the path, names six patterns and recommends starting simple.