MMarketing Against The Grain
← All frameworks
Innovation

Agent Layer Mental Model

Treat apps as data services beneath an assistant that executes user intent.

Difficulty
Advanced
Time to result
~months to results
Steps
5
Confidence
97%

The agent-layer model places an intelligent assistant above conventional software. Instead of making users open several applications, search each interface, and manually combine information, the assistant interprets intent and retrieves the required data across those applications. The underlying apps remain important because they create, hold, and update the information, but their visible interfaces may receive less direct usage. Over time, the same layer can move from retrieval into creation and execution. Product teams therefore evaluate value through data quality, interoperability, action reliability, and completed user outcomes rather than screen time alone. The model predicts both a new software architecture and a shift in incentives for application builders.

Origin

Extracted from Marketing Against The Grain while the hosts analyzed Siri retrieving flight, email, and map information across applications.

Core principles

  • 01The assistant becomes the user's primary interface.
  • 02Applications increasingly create, store, and expose data.
  • 03Agents coordinate information across application boundaries.
  • 04Retrieval precedes more autonomous action.
  • 05Software value can persist even when direct app usage falls.

How to run it

  1. 1

    Start with user intent

    Define the outcome the user wants without assuming they will open a particular application.

    Pro tip Write the intent in natural language as the user would express it.

    Watch out Do not redesign around the existing screen sequence.

  2. 2

    Map the supporting software

    Identify which applications create or hold the information needed to satisfy the intent.

    Pro tip Separate authoritative data sources from convenient secondary copies.

    Watch out Conflicting data sources require an explicit resolution rule.

  3. 3

    Create the retrieval layer

    Allow the assistant to find, combine, and summarize relevant information across applications.

    Pro tip Begin with read-only retrieval before enabling consequential actions.

    Watch out Unstructured data may require stronger confidence checks.

  4. 4

    Add bounded actions

    Permit the assistant to create or change information only in clearly defined workflows.

    Pro tip Make approvals proportional to consequence and reversibility.

    Watch out Do not give broad write access before validating retrieval accuracy.

  5. 5

    Measure outcomes

    Track whether users complete their intended tasks faster and more reliably, even if direct app usage declines.

    Pro tip Use task completion and correction rates as core metrics.

    Watch out App opens may become a misleading measure of product value.

In the wild

Coordinating an airport pickup

A user asks when their mother will arrive. The assistant retrieves flight details from an email, checks the current flight status, consults maps, and calculates the trip without requiring the user to open each application.

Several application workflows collapse into one natural-language interaction.

Preparing for a conference talk

An assistant finds the speaking time buried in email, identifies the user's current location, calculates the route, and includes a preferred coffee stop along the way.

The assistant converts scattered contextual data into a practical itinerary.

Common mistakes

Removing apps before replacing their function

Applications still supply data and domain capabilities even when their interfaces recede into the background.

Automating actions before retrieval works

Action-taking amplifies errors, so teams should first establish reliable cross-application retrieval.

Optimizing for app engagement

Direct usage can fall while the application's data and services become more valuable to assistants.

Is it for you?

Best for

It is best for teams building applications, APIs, or workflows likely to sit beneath AI assistants.

Not ideal for

It is not ideal for experiences whose value depends on immersive, direct, or highly visual interaction.

From the transcript

we talked about software getting pushed into the background because there's this agent layer on top of it.

Kieran Flanagan · 06:30

The apps are there to create the data for the assistant, but the assistant is actually the one retrieving the data on your behalf.

Kipp Bodnar · 07:00

We are moving to the agent layer. Agents are going to be the thing that sits on top of software.

Kieran Flanagan · 25:30

From the episode

Apple Intelligence Can Outsmart 99% Of Marketers (WWDC Recap)