Skip to content

AgentHerder

/ˈeɪdʒənt ˌhɜːr.dər/ (noun)a practitioner who orchestrates fleets of AI coding agents; one who makes sense of parallel, asynchronous agent work without losing the plot.

The practice of working with — not just using — AI.

Running one Claude Code session is using AI.
Running fifteen at once, on purpose, without losing control, is a practice.

We call that practice Agent Herding. This is where to learn it.

Join the certification waitlist → Cert exam opens 2026 · Free curriculum on signup

Why this is a practice, not a skill

For the last 40 years, software engineering assumed one human, at one terminal, writing code. We built our tools, our workflows, our mental models around that shape.

AI coding agents broke that shape. A single developer running 10+ Claude Code sessions in parallel is not just a faster developer. They're doing a qualitatively different job. The bottleneck moved from "how fast can I write code" to "how well can I brief, supervise, and coordinate autonomous workers."

No one taught us how to do that. There's no book. No conference talk. No training program.

Until now.

Agent Herding is the practice of:

  • Decomposing work into briefs an agent can actually execute independently.
  • Parallelising the briefs across sessions without them stepping on each other.
  • Monitoring without micromanaging — knowing when an agent is stuck, off-course, or quietly shipping.
  • Coordinating outputs back into a coherent project, without becoming a bottleneck yourself.

It looks, honestly, a lot like managing a small team of junior engineers — except the juniors cost pennies, never sleep, and produce code faster than you can read it.

Getting to 2–3 parallel sessions is usually self-taught — common sense and a weekend. The wall most people hit is the jump from 3 to 4–12, where the coordination overhead changes shape entirely. That's the step this practice is about.

Three levels of the practice

Agent Herding has three natural stages. Each one sets up the next.

Level 1 — Solo

1–3 parallel sessions. The baseline. CLAUDE.md, skills, context management, briefing a single agent well. This is common sense plus a bit of craft — most developers can learn it alone. Curriculum is here for reference, but you probably already have this.

For: reference, or if you've never run a Claude Code session before.

Level 2 — Parallel

3–8 parallel sessions. Where the practice actually begins. Decomposing work into independent streams, briefing a cohort at once, git workflows that don't collapse under merge conflicts, catching agents that are quietly off-course. This is the jump most people plateau at.

For: senior engineers, leads, and solo founders who have tried to scale beyond 3 sessions and hit a wall.

Level 3 — Fleet

10+ parallel sessions. Exam-grade. Agents spawning agents, forked sessions for alternative approaches, cross-repo coordination, and staying the coordinator instead of becoming the bottleneck. This is what the certification exam actually tests.

For: the practitioners at the frontier — and anyone preparing for the certification exam.

The certification

All curriculum is free. Email-gated just so you can pick up where you left off. The paid thing is the exam.

The exam is 60–90 minutes in a private git repo we provision for you. It contains more tasks than any single developer can complete serially in that window — the only way to pass is genuine parallelism. You bring your own fleet.

At the deadline you submit two things:

  • The final state of main — whatever your agents managed to ship.
  • A sessions/ folder containing every Claude Code transcript you used during the exam (one cctabs export-all command).

Each submission is graded by a Principal AgentHerder — someone who already runs 10+ parallel sessions daily. They read the code and the transcripts. The credential proves both that your fleet shipped, and that you knew how to coordinate it.

Join the waitlist to get:

  • Notified when the free curriculum opens.
  • Notified when the exam is live.
  • Early-bird exam pricing for the first 50.

No spam. One email per milestone. Unsubscribe with one click.

Want a structured path to get cert-ready? The AgentHerder Bootcamp is the productized 6-week prep program — async, on your real repo, outcome-guaranteed. See the curriculum →

The practice needs a tool. The tool is cctabs.

You can't herd 15 sessions from a keyboard and a pile of terminal windows. You need a tool that lets Claude open its own tabs, know which tabs are running what, and coordinate across them.

We built one. It's open source, MIT, npm install -g cctabs.

cctabs ≠ AgentHerder. The tool is the lever. The practice is what you're learning to do with it.

cctabs.com → Read the docs →

Who's behind this

Built by Fredrik Wollsén — solo practitioner at Augmented Mind, previously Mozilla. Coaches AI-native engineering practices and growth engineering, and runs 12–15 parallel Claude Code sessions daily across client work and independent products (Remember This, My Transcriber).

Everything you'll learn in this program is something the curriculum author does every day, not something they theorised about.

Currently on stage at: GOSIM AI Vision Forum Paris 2026, Panel 1 — Agentic AI Systems: Human-AI Symbiosis (May 4).

Writing

  • The AI-Native Engineering Playbook: Crawl, Walk, Run, Fly — the baseline of the practice. 1,100+ reads.
  • 14 Parallel Claude Code Sessions: A Day in the Life of an Agent Herdercoming May 2026 — what it's actually like at Level 3.
  • Subscribe to Positively Fred — new writing on Agent Herding, AI-native engineering, and the products built with this practice.

AgentHerder — the professional practice. cctabs — the tool (MIT).