How to Actually Choose the Right AI Agent

How to Actually Choose the Right AI Agent

by Nate Herk | AI Automation · View the channel on PSI

Applied Sciences & Engineering AI — News & Applications UYComputer scienceUYQArtificial intelligence (AI)
🎙 Nate Herk | AI Automation 👥 1.0M 📅 September 11, 2026 ⏱ 34 min 👁 1K 📄 expert opinion 🧭 2026-09-11
Available in: English (current) Français

Keywords

AI agentharnessmodel-agnosticskillsagentic loop

Summary

In this podcast-style video, Nate Herk and Mark Kashef discuss the crucial distinction between AI models and the harnesses that surround them, arguing that the harness—the tools, skills, and execution loops—often matters more than the model itself. They illustrate this with a diagram showing capabilities like file reading, writing, and bash access, which are not inherent to the model but provided by the harness. They compare Claude Code and Codex, describing Claude Code as a planning-oriented visionary and Codex as a precise executor, and recommend using both in a loop to maximize results. They also explore local models via LM Studio, noting that many tasks can be handled by open-source models, reserving top-tier models for complex planning. The conversation covers strategies for keeping skills portable across different harnesses, such as creating model-agnostic skills and using a dedicated conversion skill. They discuss the concept of an ‘AI operating system’ and recommend periodic audits to prune unused skills, using tools like a rot.md file to track decay. They reference the Claude Code harness leak to illustrate that models rely on numerous crutches, reinforcing the importance of the surrounding infrastructure. Ultimately, they advocate for building a personal, provider-agnostic harness to future-proof one’s workflow.

200 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides substantial value by shifting focus from model benchmarks to the practical infrastructure that determines real-world performance. The argumentation is built on concrete examples, such as demonstrating LM Studio’s inability to spin up a local server, and comparing the behaviors of Claude Code and Codex. The emphasis on making skills portable and maintaining an adaptable harness is well-reasoned and actionable. However, the claims are largely based on personal experience and observation rather than controlled experiments or data, which limits the strength of the argumentation. The discussion is coherent and persuasive for practitioners seeking practical guidance.

Scientific Rigor, Source Quality, Title Accuracy

The video lacks formal scientific citations, relying instead on experiential evidence and references to widely known tools (e.g., Claude Code, Codex, Hermes). The mention of the Claude Code harness leak is credible but unverified. The title is accurate and directly matches the content, focusing on the selection of AI agents. The presence of a sponsorship segment for Clay is clearly marked and does not interfere with the informational content. Overall, while not rigorously sourced, the video’s practical nature makes the lack of formal references acceptable for its intended audience.

201 words

Title / Content Match

The title accurately reflects the content, focusing on the process of selecting AI agents and emphasizing the importance of the harness over the model itself.

Quality & Reliability

7/10

The video offers practical, experience-based advice but relies on anecdotal evidence and personal opinions rather than formal citations. The discussion is insightful and internally consistent, but lacks empirical validation.

Chapters

Cited Sources

External References

Contribution & Novelties

The video contributes a clear articulation of the harness-model dichotomy, emphasizing that operational factors surrounding an AI model determine its practical utility. It offers a novel perspective on managing AI tools by advocating for model-agnostic skills and a personal, portable harness. The advice on periodically deleting skills and using a rot.md to manage decay provides a systematic approach to maintaining an AI setup. While many points echo common wisdom in the AI automation community, the synthesis of these ideas in a single discussion is valuable.

Pour aller plus loin :

  • AI agent — Provides a foundational definition of AI agents, useful for understanding the broader context.
  • Prompt engineering — Relevant to the skill design and portability across models discussed in the video.
  • LM Studio — The tool mentioned in the video for running local models, directly related to the local model discussion.

142 words

Radar Profile

The radar profile shows strong scores in information quantity (8) and technical level (8), indicating a technically rich and comprehensive presentation. Quality (7) and reliability (6) are moderate, reflecting the informal, experience-based nature of the content. The overall shape suggests a video that is informative and technically deep but not rigorously substantiated.

Reliability 6/10