Your AI Team Is a Slop Machine (Here's the Fix)

Summary

The speaker argues that building an effective AI team requires more than adding agents or choosing the newest model. The goal is to reduce constant human supervision while maintaining reliable, high-quality work. He presents five practices that he says apply across agent platforms such as OpenClaw, Hermes, and Grok Bot.

1. Give each agent one role. A single agent handling support, server issues, writing, and other tasks can become confused as its context grows. The speaker calls this “contextual decay”: more context does not necessarily improve performance, and semantically similar but irrelevant material retrieved through RAG can also mislead an agent. Instead, assign agents distinct responsibilities, instructions, memory, tools, and access. Clear roles also make mistakes easier to reproduce and diagnose. More agents add coordination and token costs, so the benefit is clearer boundaries—not simply a larger team.

2. Give agents distinct names. Names help the human operator remember who does what, follow handoffs, and communicate naturally, especially in voice messages. The speaker recommends names that sound different to reduce transcription errors. Names are an organizational aid for the human, not a way to make a model smarter.

3. Choose models for reliability and task fit. The speaker distinguishes model capability from dependable performance: impressive demonstrations can still require substantial monitoring and correction. He recommends using stronger models for demanding creative work and faster, cheaper models for routine checks, rather than using one model for every task. He also argues for reducing dependence on a single provider by exploring capable local or open-weight models, since pricing or usage-limit changes could affect an entire operation. He describes measuring faster GPT token consumption and using a local model for video editing as examples from his own experience.

4. Have agents review one another’s work. Agents are not reliable judges of their own output, so the speaker recommends assigning a separate reviewer with a different role. In his example, a YouTube strategist reviews scripts and editing work, catches issues such as weak introductions or overloaded graphics, and sends work back for revision. He suggests defining task owners and required reviewers, asking reviewers to support their conclusions with evidence, and keeping the human as the final reviewer for work published under their name. Different models may also provide more varied perspectives.

5. Hold structured meetings with the AI team. When a decision spans information held by several specialists, the speaker recommends collecting their perspectives in a group discussion. His example uses a Telegram group with a turn-taking process: agents first give independent opinions, then read and respond to others. A product specialist, strategist, and accountant may surface different considerations about a marketing idea. The value comes from agents with genuinely different roles and knowledge—not several copies of the same model.

Overall, the video presents AI agents as a team to be organized and managed, rather than either one all-purpose chatbot or an unsupervised swarm. Its practical guidance centers on role separation, clear ownership, cross-review, model selection, and structured coordination.

Main Speaker and Source

• Simon Høiberg, describing methods and examples from managing a small SaaS company with a team of AI agents.