I Spent $10,000 Testing AI Agents (Here’s the Stack I Actually Use)
Published
Should I watch this?
Skip to the real content 0:57Your full summary is ready below. Read on while we finish your watch recommendation.
Your summary is ready. The watch recommendation is next.
Worth watchingYes
Concrete examples of coordinating local and cloud agents, updating apps, preparing YouTube drafts, and sharing integrations make this a useful workflow tour, though it doesn't substantiate the $10,000 “best stack” claim.
The main answer
The creator’s recommended stack divides work by role rather than relying on one AI: • Claude Code, coordinated through Claude Projects: create apps, interface designs, slides, graphics, and edited videos. • Codex, coordinated through a voice agent called “Dots”: handle administration, computer/browser tasks, email, Slack, calendars, and preparing YouTube uploads for review. • Composio: centralize app connections, reusable skills, and API access so different agents can use the same tools. • “Grockbot,” as transcribed: handle Twitter/X research and account tasks, with other agents delegating to it through webhooks. The central idea is to talk to a coordinator that delegates work to local or cloud sessions. Cross-platform memory remains an unsolved part of the creator’s setup. This is a personal workflow recommendation—not a demonstrated, objectively best stack.
Key takeaways
- The creator prefers Claude for visual and creative production, and Codex for business coordination and computer use. These are subjective judgments based on their work.
- Voice is central: they describe assigning tasks, reviewing emails, and dictating replies while walking.
- Local sessions access files and applications on their computer; cloud sessions handle work that does not require that machine. Remote local work requires the computer to remain available.
- A shared integration layer can reduce repeated setup when switching agents, but shared connections and skills do not automatically provide shared memory.
- The video-editing examples required iteration. The YouTube-upload workflow saves a draft for human review rather than necessarily publishing autonomously.
- The transcript does not provide a testing methodology, comparative benchmarks, recurring costs, or failure rates to substantiate the claimed $10,000 search for the best stack.
- Broad access to email, messages, files, social accounts, and production deployments creates privacy and operational risks that the video barely discusses. Product names, availability, and capabilities are not independently verified here.
Summary
The video presents a content creator’s AI-assisted operating system: Claude makes assets, Codex handles administration, Composio connects tools across agents, and Grockbot supplies Twitter/X access. Examples emphasize delegating through voice and coordinating multiple sessions from a phone. The workflow is extensive but still depends on custom setup, iteration, and unresolved memory sharing.
What’s being promoted?
Explicit paid promotion: Superhuman sponsors a clearly identified segment. The creator promotes its Mail and Docs products through a workflow that turns email deadlines and feedback into a content calendar, then directs viewers to a description link. They say they already use it. The transcript does not establish whether that link also earns affiliate commissions. Own-product promotion: The creator introduces V2, gives the address “v2.vide,” and asks viewers to join an email waitlist. They describe it as a forthcoming free, open-source, editor-agnostic version-control platform for AI video workflows. This is a clear audience-building and potential customer-acquisition purpose alongside the tutorial. Availability, licensing, and the promise that it will always be free are not demonstrated. Strong recommendation without disclosed sponsorship: Composio receives an extended demonstration and repeated praise. The creator explicitly says twice that it is not sponsored. That supports treating it as an enthusiastic recommendation, not a confirmed paid placement; the transcript cannot establish whether any other commercial relationship exists. Broader business positioning: The creator mentions starting a company called Agent Native. Demonstrating sophisticated workflows also builds their reputation and attracts an audience interested in agent-native products. The $10,000 testing claim, follower count, and video-engagement figures reinforce authority, but are not substantiated within the transcript. Native Note mainly serves as an internal-tool demonstration; other named services mostly provide workflow context. Overall: the video combines practical instruction, a disclosed advertisement, self-promotion, and personal-brand building. Those incentives may encourage enthusiastic framing, but there is no basis in the transcript to conclude that the recommendations are dishonest or secretly sponsored.
