Walk Before You Agent
Master predictable automation before adding agents and AI decisions
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 98%
Begin with a recurring task rather than an ambitious autonomous agent. Observe the task, list its steps in order, and identify the data source, transformation, and destination at each stage. Build the simplest deterministic workflow that transfers this data and performs the known actions. Only introduce AI when a step genuinely requires interpretation, generation, or judgment that fixed logic cannot provide. This sequencing teaches transferable concepts such as variables, JSON, and left-to-right data movement without adding agentic complexity too early. The result is a predictable automation that produces an immediate win, exposes where intelligence is actually useful, and provides a stable foundation that can later be expanded into an agent.
Origin
Nate Hurk described this approach on Marketing Against the Grain while explaining how nontechnical beginners should choose and build their first automation.
Core principles
- 01Automate a stable process before introducing an agent
- 02Use AI only where judgment or interpretation is necessary
- 03Repeated ordered steps signal a strong automation opportunity
- 04Tool-specific knowledge matters less than workflow fundamentals
How to run it
- 1
Find a recurring task
Choose something performed daily or weekly. Frequent repetition creates enough value to justify automation and provides multiple runs for validation.
Pro tip Review existing SOPs because they often reveal high-frequency processes with documented steps.
Watch out Do not start with an agent merely because agents appear more advanced.
- 2
Map the process
Write every action in the order it is normally performed. Record who acts, which data source is used, and what transformation occurs.
Pro tip A process that reliably follows the same sequence is an especially strong candidate.
Watch out Do not automate a process whose actual execution differs substantially from its documentation.
- 3
Build the deterministic path
Connect the systems and move data through the fixed steps without AI wherever possible. Confirm that the workflow follows the same path on every run.
Pro tip Start with a single trigger, one transformation, and one destination.
Watch out Unnecessary branching makes early workflows harder to diagnose.
- 4
Locate judgment points
Identify steps that cannot be completed with explicit rules, such as interpreting free text or generating customized content. These are the candidate locations for AI.
Pro tip Ask whether a conventional rule could handle the step before selecting a model.
Watch out AI introduces nondeterminism, cost, and new failure modes.
- 5
Add and constrain AI
Insert AI only at the selected judgment points and give it narrow inputs and outputs. Test normal cases and likely exceptions before increasing scope.
Pro tip Keep the surrounding workflow deterministic so model behavior is easier to inspect.
Watch out Do not grant broad autonomy before the narrow version is reliable.
In the wild
A hypothetical operations team copies form submissions into a spreadsheet, calculates fixed fields, and emails the result every Friday. They first automate the transfer, calculations, and delivery without AI. They later add a narrowly scoped model only to summarize free-text comments.
→ The team removes repetitive work while keeping calculations and delivery predictable.
A hypothetical agency selects a ten-step client-onboarding SOP, labels each data source and owner, and automates the fixed account-creation and notification steps. Human approval remains in place for exceptions until those cases are understood.
→ The agency gets a reliable first automation without prematurely building an autonomous agent.
Common mistakes
Starting with a complex agent
A beginner may combine tools, memory, branching, and autonomous decisions before learning how data moves through a basic workflow. The resulting system is difficult to debug and trust.
Using AI for fixed logic
A model should not perform a transformation that ordinary workflow logic can complete reliably. Doing so creates needless variability and expense.
Choosing tools before processes
Debating platforms can displace the more important work of identifying repeatable tasks and understanding workflow fundamentals.
Is it for you?
Best for
It is best for teams beginning to automate recurring operational tasks with N8N, Make, Zapier, or similar tools.
Not ideal for
It is not ideal for inherently exploratory work whose steps and desired outputs change on every run.
From the transcript
“don't run before you walk.”
“And if I do those three steps in that order every single time, that is a great great use case for something that can be…”
“my mindset is how can I eliminate as much AI as possible.”
From the episode
This Ai Agent Turns 1 Image Into A 30 Second Commercial