Pareto 20, No. 14: Cole Medin, for building agents that hold up in production

He spent millions of tokens testing a benchmark claim and did not buy it. Code-first, sponsor-funded, and one of the few automation educators who tells you when a tool is not ready.

Seat 14 on The Pareto 20 is Cole Medin, and it is the seat for the operator who has moved past a visual automation builder and is now asking an agent to do real work.

Who he is

Medin is a developer and educator in Minnesota who left a Fortune 500 application-developer job in October 2024 to found Dynamous, an AI education company, and became CTO of oTTomator, an open agent community. He builds mostly in Python with Pydantic AI and Claude Code, maintains widely used open-source projects including Archon (past 22,000 GitHub stars in July 2026), and publishes on YouTube and LinkedIn several times a week. In August 2026 he wrote that he had not typed a line of code himself in over a year: everything goes through agents, which is the experiment his content documents.

What to read him for

A benchmark claim, tested and rejected. On July 24, 2026 he took the claim that an open-weight model had caught up with the frontier and ran it through deliberate failure traps: false premises, hidden rules, flattery bait. His result: the challenger failed 36% of the traps against 8% for the incumbent. "I spent millions of tokens testing that claim, and I'm not buying it." His general rule from the same post: "Benchmarks are mostly clean, well-scoped tasks the models are practically trained on."

When a promising tool is not ready. On DeepSeek's coding agent harness (August 20, 2026): "it's also rough around the edges... if you want something that just works today, those are still the better call." That is the sentence most tutorial channels never write.

The setting you should not flip. On the permission-skipping mode every coding agent now ships (July 23, 2026): "Every coding agent ships a YOLO mode now, and it's genuinely dangerous." His fix is sandboxing rather than better prompts, which puts him in agreement with Simon Willison and Laurent DeSegur from a different starting point.

Rules files that decay. August 6, 2026, on over-long instruction files for coding agents: "Rules that fix reasoning have decayed, because the models got good at reasoning. Rules that direct attention have not." Useful if your agent's instruction file has grown past what anyone reads.

Against switching tools every week. His September 1, 2026 video pushes back on the churn narrative and argues for improving the setup you have. Coming from someone whose income depends on new content, that is a verdict against his own incentives.

Start here

The channel, most recent first. His July 16, 2026 walkthrough of an open-source agent framework (video) shows the production concerns he cares about: durability, sandboxing, approval gates.

What he sells, and what that means for you

Dynamous membership at $72 a month or $712 a year (discounted from $80 and $949 at the time of checking), which bundles three courses and weekly live sessions; consulting with companies rolling out coding agents; and video sponsorships that in the last year included Lindy, Sonar and AWS. Lindy is a no-code agent builder and Sonar a code-quality tool, both in categories he covers, so a warm mention of either should be read with the sponsorship in mind. Archon is free and open source. His marketing addresses "innovators and founders" alongside developers; in practice the content is written for people comfortable with Python, and a no-code operator will hit that wall quickly.

Not for you if

You are not going to touch code and do not have someone who will. This is the developer end of the automation domain; Nate Herk and Dheeraj Sharma hold the other end.

---

Part of The Pareto 20, twenty creators who get a solo operator most of the way on AI decisions. How we chose them. Facts above were checked against the creator's own pages on September 12, 2026.