What Is Agent Herder? Supervising AI Agents in the Terminal

Table of Contents

• The specific problem it solves
• What "Agent Herder" actually is in 2026

  • Watch out for the deprecated package
• Real requirements (read this before installing)
• How to install it
• When Agent Herder is the right call
• When not to use it
• Herder is not a harness — it is a layer on top of one
• Conclusion
• Keep reading

The specific problem it solves

With a single agent, the chat is enough supervision: you see what it does, approve or correct, move on. That model breaks the moment you run work in parallel. One agent refactors a module, another writes tests, a third investigates a bug. Without a coordination layer you end up alt-tabbing between terminals, losing track of which session is blocked and which already finished.

The classic answer is a terminal multiplexer like tmux. It works, but it forces you to memorize shortcuts and to be the one who remembers what lives in each pane.

Agent Herder inverts that. Instead of you managing the panes, Claude itself manages its fleet of sibling sessions: opening them, reading their output, sending them prompts, and forking a conversation to explore an alternative — without you switching tabs by hand. You supervise; the agent coordinates.

What "Agent Herder" actually is in 2026

There is a confusion worth clearing up here, because it changes what you install.

"Agent Herder" today is not a single package but a small ecosystem with two pieces you should not mix up:

1. The tool. The binary that actually orchestrates, shipped as @generativereality/agentherder. Open source and free.

2. The education layer. agentherder.com operates as a community and training program around the practice of coordinating agent fleets: a free self-study curriculum, a multi-week bootcamp with mentoring, and a paid certification exam.

That second piece is distinct from the tool and aims at teams that want to standardize how they work with multiple sessions, not just install a CLI. If you want the tool, you do not need to pay for anything.

Watch out for the deprecated package

The original package, @generativereality/herd, is marked as deprecated on npm in favor of its successor. Worth knowing if your starting point is an old link: installing the old one today redirects you to the active version, but it is better to go straight to @generativereality/agentherder and not get stuck on an unmaintained release.

Real requirements (read this before installing)

This is the part most people discover too late. Agent Herder has platform constraints that are not a minor detail:

Requirement	Detail
Terminal	Wave Terminal required. It does not run on any arbitrary terminal.
Operating system	macOS only, with accessibility permissions granted to Wave.
Runtime	Node.js 20 or later.
Agent	Primarily designed for Claude Code.
Interface	Terminal tabs only. No GUI, no mobile app.

If your team runs Linux in CI or WSL on Windows, this tool simply does not apply today. It is not a limitation you can configure around: it is a platform dependency.

How to install it

There are three routes, all equivalent in outcome:

# 1. Global npm package
npm install -g @generativereality/agentherder

The other two are installing it as a Claude Code marketplace plugin or as a standalone skill, depending on how you want it wired into your workflow. If you already manage your Claude Code plugins through the marketplace, that route keeps everything in one place.

Before running any of the three, make sure Wave Terminal is installed with accessibility permissions, and Node 20+ is available (node -v). Order matters: installing the package without the host terminal leaves you with a CLI that has nowhere to open tabs.

When Agent Herder is the right call

The scenario where it fits is fairly specific, and that is a virtue, not a flaw:

• You are a solo developer or part of a small team.
• You already live on macOS.
• You already use Wave Terminal, or you are willing to adopt it.
• You work almost exclusively with Claude Code.
• Your actual workflow is "spin up three or four sessions that coordinate with each other to split one large task".

In that case the named-tabs model is lighter than standing up a full development environment: no worktrees to learn, no embedded browser, no coordinator concepts. Just tabs that Claude opens and reads itself.

When not to use it

Be honest with yourself here, because the installer will not warn you:

• You work on Windows or Linux → not available.
• You need to supervise from your phone → there is no app.
• Your fleet spans several providers (Codex, Gemini, Cursor) → it is built for Claude Code.
• You want GitHub or Linear integrations → it ships none natively.

In all four cases, the reasonable alternative is Orca, which is cross-platform, multi-agent, and has a mobile app in beta.

Herder is not a harness — it is a layer on top of one

A conceptual note that prevents a common misunderstanding: searching "herder ai harness" conflates two things. The harness is the environment that gives the model its tools, its context, and its permissions — that is Claude Code here. Herder does not replace that harness: it multiplies it, letting several instances of the same harness coordinate.

If what interests you is why that layer matters more than picking the model, we cover that separately in why the harness matters more than the model.

Conclusion

Agent Herder is a deliberately narrow tool: terminal tabs on macOS so Claude Code can coordinate its own sessions. Inside that scope, the learning curve is low and trying it costs nothing. Outside that scope — different OS, different agent, remote supervision — compatibility is not partial, it is zero.

The useful question before installing is not whether it is a good tool, but whether your workflow is already 100% terminal and 100% Claude Code on macOS. If yes, it weighs less than any alternative. If no, do not force it.

Keep reading

• Full comparison: Orca vs Herder: which orchestrator to choose
• The cross-platform alternative: What Orca is, the ADE for parallel agents
• Context: Why the harness matters more than the model