Most Valuable Skill of 2026: Managing AI Agents

Guest: Ryan Carson · Host: Greg Isenberg | Published: 2026-07-24 | Source: transcript (44:47, 160K views at ingest)

Summary

Greg Isenberg has Ryan Carson (25-year founder/CEO; Treehouse acquired, ~110 employees; now solo at Untangle, an AI divorce/paralegal agent for family-law firms, seed-stage, claiming a near-term 4× revenue month) walk through how he runs a team of AI employees. Everyone — IC, manager, founder, parent, student — is now a manager of agents; world-class operators will outperform everyone else. Three skills: cloud agents, automations, a software factory. The stack changes every 3–6 months; reps beat theory.

Desk: eight panes on a 52” monitor — Slack, 1Password (never give agents prod-write keys; they ask, he pastes), Devin (Cognition) as the factory, two test panes, X, Codex for local computer-use. Voice via Wispr Flow; analog Ugmonk list for the day’s important ships. ~50%+ of work is iPhone Devin, anywhere a PR is blocked on a human yes.

Load-bearing claim: cloud, not local. Local Mac + git worktrees cannot run 5–10 parallel agents without collisions; Devin (and increasingly Cursor/Codex) spins an isolated VM per session. Carson often runs 5–10. Local-only is “caveman” / “shipping 10× less”; X “local maxis” are “not doing real work.” Carve-out: heavy new UI starts local, then moves to cloud. Output: ~22–25 PRs/day, sometimes 40. The job is no longer 2–3 high-stakes decisions a day; it is 10–20 by lunch. Pin big threads, let fire-PRs rip, check ~every 25 minutes, accept you will work more not less. A parent-model thread (spoken “Fable”) spins cheaper children (“Fusion”) — you still cannot skip being the manager.

Automations: have the agent build the loop. (1) A Devin playbook (checklist, not a skill) browser-tests signup→discovery 3×/week (~$60 tokens), records video, self-fixes, child-triages, posts to Slack via MCP. (2) A 9am production watchdog rolls paid-customer events into admin JSON with links into real UI. (3) A daily self-improvement loop for in-product agent Grace: grade chats on a rubric, spawn a child PR if below score — ~3 paper-cut UX PRs/day. He still dogfoods the app; models “lack this sense of obvious intelligence.”

Cost: ~~$20k tokens last month (“not viable”); **~~$5k/employee/month** is the expected shakeout. Run loops on cheap fine-tunes (Cognition SWE-1.7, ~$5/session). Do not build the company factory on Claude Code or Codex (subsidized lock-in). Personal Codex at $200/mo is fine; company motion belongs on an independent agent lab (Devin, AMP, Factory). Don’t DIY a factory until Ramp-scale. Destination: agents write, review, and ship 100% of the code. Third meta-skill: share what you learn in public on X.

Key Claims

Everyone is now a manager of agents; world-class agent operation is the 2026 skill. Becoming a better agent manager makes you more technical, not less. Engineering is not going away. (Carson)
Work in the cloud, not locally — isolated VMs let you run 5–10 agents without worktree collisions. Local-only is “caveman” / 10× slower. Heavy new UI is the exception, then move to cloud. (creator-opinion, strongly stated)
Output: ~22–25 PRs/day (sometimes 40); ~50%+ of work from phone. Job = 10–20 high-stakes decisions by lunch. Pin big threads; check ~every 25 minutes. You will work more, not less. (self-reported)
Never give agents prod-write keys; 1Password + paste-into-session. (practical)
Automations to build: scheduled browser-test playbook, production watchdog (customer events → linked UI summary), self-improvement rubric on the in-product agent (Grace) → child PRs. Have the agent design the machinery (Slack MCP, child triage).
Humans must still use their own product; models lack “obvious” taste. (limit)
Token reality: ~$20k last month too high; ~$5k/employee/month is the shakeout; route loops to cheap fine-tunes (~$5/session). (self-reported)
Don’t build the company software factory on Claude Code/Codex (subsidized lock-in). Use independent labs (Devin/AMP/Factory). Don’t DIY a factory until you’re Ramp-sized. Agents will write/review/ship 100% of code. (hot take + commercial)

Notable quotes

“You are a manager of agents now. So no matter what you used to do, whether it was a people manager, an IC, you are going to become a manager of agents now and you need to be the best in the world.”

Why citable: The episode’s title claim — process, not SEO — and the wiki’s cleanest one-line definition of the agent-ops job.

“If you are working locally, I honestly think you are a caveman. Like I think you are holding yourself back and you are shipping 10x less than you could be.”

Why citable: The sharpest (and most contested) operational claim in the interview; the fork vs. local Claude Code workflows.

“Your job is going to be managing as many agents as you possibly can… It is a farce that engineering is going away. If anybody tells you engineers are going away or becoming less technical, they’re just not doing the work.”

Why citable: Pairs the “manage more agents” thesis with the counter-narrative that agent management increases technical depth — useful against both job-replacement panic and “vibe code, skip the fundamentals.”

Connections

Entities mentioned: Claude Code, Claude, ChatGPT · NEW Greg Isenberg, NEW ryan-carson · Devin/Cognition, Codex, Cursor, AMP, Factory, Wispr Flow, Slack, Untangle, Treehouse, Ramp Inspect (minor) Concepts referenced: AI-Assisted Development (verification, parallel agents, augmentation vs. automation, human-validation of prod) · NEW Agent Orchestration (cloud VMs, playbooks vs. skills, parent/child sessions, scheduled loops, model routing) · CI/CD & Automation (DORA Elite) (browser-test playbooks, triage PRs, watchdog)

Contradictions / Tensions

Team-of-agents vs. god-agent: this is the opposite pole from The AI Agent Every Company is About to Build | Vercel CEO Guillermo Rauch. Carson wants 5–10 parallel cloud agents the human juggles; Rauch hides multiplicity behind one Slack @V. Same problem (how companies run agents), incompatible UX.
Cloud vs. local: Carson dunks on local; Paste This Into Claude, It'll Make You Build 10x Faster (6 Claude Code power phrases) and much of AI-Assisted Development is local Claude Code + sub-agents. Record as a real tooling fork, not a settled best practice. Carson himself still uses Codex locally for computer-use.
vs. Why Your AI Offer Isn't Selling, and How to Fix That: both want model routing and warn that subsidized frontier tokens hide unit economics. Jones: DIY auto-router, 95% open models. Carson: pay an independent lab to route for you. Jones says jobs survive because jobs ≠ tasks; Carson says engineering survives because managers of robot-carpenters still need to know where the nail goes.
vs. Marchese on automation: Marchese prefers augmentation over automation (operational debt); Carson schedules expensive browser loops and self-improvement PRs. Different stage (personal coding agent vs. PMF SaaS with real law-firm customers).
Commercial bias: Devin is not cheap and is the hero tool; $20k / 22–40 PRs / 4× revenue are unaudited. “Cursor owned by Elon” is almost certainly wrong. “Claude Code/Codex will lock you in” is incentive-talk from a Cognition customer.

Notes

Process source, not SEO — but it is the wiki’s most concrete agent-ops runbook (keys, cadence, playbooks, watchdog, rubric loop, token budget). Strongest evidence so far for a dedicated Agent Orchestration concept. NEW entities later: greg-isenberg, ryan-carson (and possibly devin / Cognition if the cluster grows). Backfill: Cognition SWE-1.x model names; Carson’s Untangle/Treehouse claims; whether “Fable/Fusion” are Cognition product names or ASR.