Agent-or-App Complexity Threshold
Use agents for narrow processes and apps for complex, data-rich systems.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 98%
This threshold uses workflow complexity to choose between an agent and an application. A narrow agent is appropriate when a specific, repeatable process has standard inputs and a clear output. An application becomes justified when the solution must combine several data sources, coordinate multiple functions, retain state, expose an interactive interface, or execute logic too complex for a single agent workflow. The decision should begin with the business problem rather than a desire to build software. Teams first test the task with prompts and agents, then graduate only when those simpler forms become structurally inadequate. Even at the application level, the preferred scope is a focused point solution addressing a specific friction in the customer journey.
Origin
Extracted from Marketing Against The Grain during a comparison of narrow agents with AI-coded applications such as workout tools and a marketing index.
Core principles
- 01Choose the simplest architecture that solves the problem.
- 02Use agents to automate specific processes.
- 03Graduate to an app when data sources or interactions multiply.
- 04Tie applications to concrete customer-journey problems.
- 05Let complexity, not novelty, justify application development.
How to run it
- 1
Define the Use Case
Specify the user, problem, inputs, desired result, and frequency. Remove features that are unrelated to the central outcome.
Pro tip Frame the first version as a point solution.
Watch out A vague aspiration will produce unnecessary architecture.
- 2
Try the Simplest Form
Determine whether a prompt or narrow agent can already solve the process. Prototype the core logic before selecting an application stack.
Pro tip Use the prompt prototype to clarify requirements.
Watch out Do not assume a graphical interface is necessary.
- 3
Assess Complexity Signals
List data sources, APIs, interacting workflows, state, algorithms, user roles, and interface needs. Identify which elements cannot be handled cleanly by a narrow agent.
Pro tip Multiple independent complexity signals provide stronger justification than one optional feature.
Watch out One API call alone does not necessarily require an app.
- 4
Choose Agent or App
Use an agent for a specific repeatable process with bounded inputs and outputs. Choose an app when the solution must orchestrate substantial data and functionality.
Watch out Avoid combining several unstable agents merely to postpone a justified application design.
- 5
Build the Point Solution
Implement only the functions needed to solve the defined problem. Use AI coding tools to accelerate development while reviewing architecture and security choices.
Pro tip Require explicit handling for credentials, errors, and data quality.
Watch out Do not blindly accept every technology or deployment decision selected by the coding agent.
- 6
Validate the Value
Test whether the solution improves the targeted workflow and whether the added complexity is sustainable. Expand only after the core result is proven.
Pro tip Compare against the original manual or agent-based process.
Watch out Technical completion does not prove business usefulness.
In the wild
The hosts use Replit's coding agent to build a focused interval-training application in a few minutes. The app is appropriate because the desired experience includes an interactive tool rather than a single generated answer.
→ A non-specialist can turn a specific idea into a usable point solution.
An application combines an S&P 500 company list, marketing data from public APIs, a scoring algorithm, company rankings, and historical growth analysis. The multiple data sources and analytical functions exceed a narrow content agent.
→ The builder can investigate whether stronger marketing scores relate to company performance over time.
Common mistakes
Building an App Too Early
An application adds engineering and maintenance burden when a prompt or agent may already solve the task.
Letting Tools Choose the Problem
Starting with a fashionable coding agent rather than a customer or workflow need encourages unnecessary products.
Ignoring Technical Decisions
AI-selected languages, hosting, APIs, and deployment patterns still require human review for cost, security, and maintainability.
Is it for you?
Best for
Builders evaluating AI solutions that may involve automation, multiple APIs, algorithms, or interactive user workflows.
Not ideal for
Teams selecting technology before defining the underlying business problem and operating process.
From the transcript
“agents are really good at automating like very specific processes.”
“You kind of graduate to needing an app when you need lots of different data sources. What you're trying to do gets very complex.”
“we're gonna find areas in our customer journey and be able to build point solutions to solve those problems.”
From the episode
From Beginner to AI Expert in 30 Minutes: The 5-Step Framework You Need