Which AI agent should do which job? In this walkthrough, Riley Brown uses Claude Code for app and media creation, Codex and Dots for coordination and business tasks, Composio for app connections, and GrokBot for X-centered research. The useful lesson is the division of labor, not the number of subscriptions. Riley says he spent $10,000 testing tools; that is his reported research spend, not a setup cost or a target for readers.
Start with the work that repeatedly stalls: building an internal tool, preparing content, or handling routine admin. Choose one agent that can do that job with your approval. Add a shared connection layer only when you have a real handoff problem; keep a written project brief because shared tools are not the same as shared memory.
Watch Riley Brown's Stack Walkthrough
Source note, 29 October 2026: The video and supplied transcript show Riley's own workflow. I checked product boundaries against the official documentation linked below, but did not reproduce his app, video edits, connected accounts, costs or follower results. The video includes a Superhuman sponsorship segment; it is not evidence that the sponsored tool improves this stack.
The Stack at a Glance
| Layer | Riley's choice | What it does in his demo | Important limit |
|---|---|---|---|
| Creative worker | Claude Code | Builds his Native Note app, makes slides and edits video. | Review code, rights and final media; local work still uses a model service. |
| Project coordinator | Claude Projects | Routes and tracks local and cloud Claude tasks. | The redesigned Projects experience is a staged beta for selected Pro and Max users. |
| Operations worker | Codex | Builds, researches and prepares business content. | Computer actions and publishing still need appropriate access and review. |
| Voice front door | Dot | Takes spoken requests and directs Codex work. | Plan, region and rollout restrictions; local access must be enabled. |
| Connection layer | Composio | Shares connected app tools across agent setups. | Scopes and spending need configuration; this is not universal memory. |
| Specialist | GrokBot | Researches X and receives work through a webhook. | Verify findings and protect the webhook secret. |
Claude: Build and Create
Riley's most concrete example is Native Note, an internal app he built for his own work rather than a general-purpose product benchmark. He also shows Claude producing presentation material and video edits. That makes Claude Code his creative production environment, not simply an autocomplete tool. For video, a draft still needs a viewing pass: inspect cuts, captions, claims, music rights and export quality before anyone publishes it.
His coordinator is the redesigned Claude Projects experience. One project can send different work into threads; in the demo a clip task runs on his computer while a reference artifact runs in the cloud. Anthropic's Projects documentation says this version is a staged beta for selected Pro and Max users, not a feature every Claude account has. Cloud Claude Code runs in a separate environment; a cloud task does not automatically see uncommitted local files. Riley can steer local work from his phone while his connected computer is on. That is different from an independent cloud task continuing after his computer is off.
Codex and Dots: Coordinate the Work
Riley uses Codex for business operations and a Dot as a voice-accessible coordinator. On a walk he asks for an app change and a research page. Later he shows computer use preparing a YouTube draft from a video and sponsor email. The transcript shows preparation, not proof that a production deployment or publication was completed. He says he keeps the final publish decision himself.
OpenAI's Dot guide says access depends on plan, region and rollout. As checked on 29 October 2026, ordinary Pro rollout excludes the European Economic Area, Switzerland and the UK; eligible business or enterprise access follows separate workspace rules. That matters for readers in Portugal. A Dot has its own cloud computer, but access to your computer is optional, off by default, and requires explicit permission in the desktop app. Local tasks depend on that connection and a reachable computer; a cloud Codex task needs a configured cloud environment. Neither route should be described as automatically private just because a file began locally.
Riley's rule of thumb is personal: Claude for visual and media work, Codex for operating the business. You can test the same job in either tool before adopting his preference. A practical comparison is a single real task, the same inputs, a time limit, a review checklist and the total cost including repairs.
Composio: Reuse Connections, Not Memories
Riley connects business apps and makes their tools available to multiple agents through Composio. His companion guide gives the setup from his perspective. Composio's session documentation explains the actual boundary: a session ties a user, connected accounts, authentication and tool access together. The catalog is discoverable by default, so restrict toolkits and actions to the task. Reusing one connection reduces repeated setup, but it also concentrates permissions in one place. Begin with a single read-only connection, then require approval before sending messages, changing files or posting.
Riley also mentions Composio Instant, which can run eligible tools on Composio's provider accounts and bill a prepaid balance. It is experimental, off by default, and has no built-in per-session dollar cap; the docs advise application-side budgets or per-call approval. A common app connection, a reusable skill and a shared project memory are three different things. Composio can help with the first two. Riley says the third remains unfinished in his setup.
GrokBot: Bring X Research Into the Loop
GrokBot is Riley's specialist for work involving X. He describes handing tasks from Claude Code or Codex to GrokBot and getting research back. His webhook guide shows the pattern; xAI's official webhook documentation requires a signed request and a private endpoint secret. Do not put that secret in a public repo or page. X posts can be useful leads for research, but an agent summary still needs source, date and context checks before it becomes a claim in your own content.
The Missing Layer Is a Reliable Handoff
Riley is candid that his agents do not yet share a dependable memory. He tried a separate memory tool but says it did not update often enough. The editorial implication is simple: tool access is not continuity. An agent may open the same app yet know nothing about what another agent decided yesterday.
Before adding another memory product, keep a small human-readable project brief: current goal, approved constraints, source links, decisions already made, next action, owner and date. Give that brief to whichever agent takes the next turn and ask it to update only after you review the result. This is an original operating suggestion, not a workflow Riley demonstrated in the video.
What Should You Try First?
| Your bottleneck | Smallest useful test | What to check |
|---|---|---|
| An internal app or visual asset keeps waiting on a team | Give Claude Code one small feature or one owned video clip. | Does the reviewed result save time after corrections? |
| Research and admin never reach a finished draft | Give Codex one sourced brief or a draft publishing package. | Can a human approve it without redoing the work? |
| Switching agents means reconnecting the same app | Connect one low-risk, read-only app through Composio. | Are scopes clear and is the handoff actually faster? |
| X matters to your work | Ask GrokBot for a cited research note, not an autonomous post. | Can you verify each useful source? |
Only expand the stack when the measured bottleneck moves. The first win is a task that reaches a better reviewed output, not another integration icon.
Video Chapters
Claude projects and production
| Time | Topic |
|---|---|
| 00:00 | Why Riley built his agent stack |
| 01:01 | Coordinate local and cloud work with Claude Projects |
| 05:53 | Build an internal app with Claude Code |
| 08:45 | Design, slides and visual assets |
| 11:30 | Edit video and inspect the result |
Codex, Dot and operations
| Time | Topic |
|---|---|
| 15:45 | Use voice to direct Codex from a walk |
| 18:57 | Prepare a YouTube draft with computer use |
| 20:11 | How Riley splits Claude and Codex |
| 21:21 | Email and Slack triage |
Connections and remaining gaps
| Time | Topic |
|---|---|
| 22:27 | Composio as a shared connection layer |
| 25:28 | Composio Instant and usage costs |
| 26:57 | The unresolved memory problem |
| 28:02 | GrokBot and the final stack |
Sources and Limits
Checked 29 October 2026. The video, supplied transcript and Riley's resource library describe his stack and subjective preferences. For product limits, see Anthropic Projects, OpenAI Dots, Composio sessions, Composio Instant and xAI webhooks. Availability, pricing and permissions can change. We did not reproduce Riley's workflows or independently audit his reported spend, performance or results.