7 Layers of the AI Agent Stack One Creator Uses After $10,000 in Testing

Video: Here Is My ENTIRE AI Setup (Nothing Held Back)

Oct 29, 2026 · based on the channel Riley Brown

The AI agent stack is a layered setup where each layer does one job: a model for reasoning, a coding harness for building, a connector layer for shared tools, and a memory layer for context. After spending $10,000 testing agent tools, creator Riley Brown settled on a core of Claude, Codex, Dots, GrokBot and Composio, split by task rather than by brand. You can test the email and scheduling side of that workflow through Superhuman's signup, which sponsors his setup.

What the AI agent stack actually contains

The stack is not one product but seven distinct layers, each with a different failure mode. Brown's breakdown maps them clearly:

1. Reasoning and writing model — Claude, used through Projects for long-running context.
2. Coding harness — Claude Code for building apps end to end.
3. Design and slide generation — Claude for decks and visual assets.
4. Video editing — Claude-driven workflows for cutting and assembling footage.
5. Voice and computer use — Dots as a voice assistant, Codex for operating a computer.
6. Connector layer — Composio, which centralizes skills and plugins so agents can share them.
7. Memory — the unsolved layer, where context still breaks when you switch tools.

That structure matters more than any single tool. A stack is only as good as the handoffs between layers, and most of the $10,000 went into discovering where those handoffs fail.

How Claude Code, Codex and GrokBot split the work

The split is task-based: Claude Code builds, Codex operates, GrokBot fills the gaps. Brown uses Claude Code as his primary app-building environment, where a single session can take an idea to a working prototype without switching tools. Codex handles computer use — driving a browser or desktop interface — and background jobs where you want the agent working while you do something else. GrokBot sits at the edge for tasks that need a different model's behavior or a quick second opinion.

The practical rule is to assign each agent a lane and stop asking one model to do everything. When two agents overlap on the same task, you pay twice and get inconsistent output. Viewers of the breakdown pushed back on exactly this point, asking whether the separation survives contact with a connector layer like Composio — the honest answer is that the lanes blur unless you define them explicitly and check which agent produced which artifact.

Composio and the shared skills problem

Composio solves tool-sharing, not memory-sharing. It lets agents share skills and plugins from one central place, so a capability you build once is available across Claude, Codex and other harnesses instead of being rebuilt per model. Composio Instant speeds up connecting new tools to that pool.

The limit is structural. Skills and plugins travel between agents, but memory does not — switching models is like moving to a new phone that keeps the apps and loses the photos. Experienced users in the discussion argue that shared memory needs a real harness-level fix: a shared folder of short, topic-based Markdown files backed by a searchable index, with hooks that inject a table of contents into the prompt at session start and write updates back at session end. That is more engineering than most creators want, but it explains why a plugin alone will not close the gap.

Where the money actually goes

Cost control is the layer most people skip, and then pay for twice. Brown's $10,000 figure is a testing budget, not a subscription bill — it covers trials, failed experiments and tools that looked brilliant in a demo and did the same job badly in production. One team quoted in the discussion let a client trial five agent tools simultaneously; every demo impressed, and three of them duplicated work.

The working method that came out of that experience:

• Test one stack against one real job, not five tools against a demo.
• Keep the tool that survives a bad week, not the one with the prettiest interface.
• Audit overlapping subscriptions quarterly and cancel anything that has not produced an artifact in 30 days.
• Budget for the connector layer and memory work, which are usually underestimated relative to model costs.

For a solo creator, the realistic entry point is two paid tools — one model with a coding harness and one connector — before adding voice or computer-use agents.

Common questions and doubts from practitioners

The recurring doubts are about memory, YouTube transcripts and where to start. Readers repeatedly ask how to get video transcripts into an agent without paying for extra API calls; the workaround is an in-app browser pulling the transcript rather than a connector that cannot access it. Others ask whether note-taking tools like Obsidian solve the memory problem — they help as a human-readable store, but only if an agent can query them programmatically.

Beginners want a map of which parts of life an agent can actually run. The honest answer from the stack is narrower than the marketing: email triage, scheduling, drafting, slide generation, video assembly and small app builds are reliable today; anything requiring persistent cross-tool memory still needs manual glue. A GitHub repository that every agent uploads updates to is a cheap, popular substitute for shared memory, though it assumes more technical comfort than most non-developers have.

Tools and resources from the setup

All links from the breakdown, unmodified:

• Superhuman signup — the email and scheduling tool sponsoring the workflow.
• https://agentnative.inc/resources — Brown's collection of workflows and resources for the stack described above.

The bottom line on building an AI agent stack

The AI agent stack that survives real work is small, layered and ruthlessly pruned. Claude Code, Codex, Dots, GrokBot and Composio each earn their place by doing one job well, while memory remains the open problem that no plugin has solved. Start with one model and one connector, test against a real task for a full week, and only then add voice or computer-use layers. Riley Brown's full breakdown walks through each layer in his video, and his Superhuman signup link is the fastest way to test the email side of the workflow.

Source: Here Is My ENTIRE AI Setup (Nothing Held Back) · Markdown version