Glossary
The herd, in plain English
Every page on this site assumes you already know what an agent, a harness, or a workflow stage is. This one doesn't.
Who this is for
You don't need to already speak this language
If you've got an idea for something to build and no engineer on staff, or you're the product person in the room while everyone else nods along to "capability routing," this page is for you. Each term below gets a plain answer to four questions: what it is, why it matters to whoever's paying for it, how it actually works in CodeHerder, and what people commonly get wrong about it.
Already comfortable with the vocabulary? Skip straight to how it works for the full platform walk-through, or the docs for the reference detail.
24 terms
Pick a term
The work
Task
One unit of work filed in plain language, the thing an agent actually gets assigned to build.
Workflow stage (stage)
One position a task passes through on its way to done, such as plan, code, review, merge, or verify.
Human in the loop
A point in a task's workflow where the decision belongs to a person, and only a person, before the work moves forward.
Agent session (session)
One running instance of a harness working a task on a device, the actual process you can watch in real time.
Separation of duties
Requiring that the person who reviews or merges a task isn't the same person, or agent, who wrote it.
Vibe coding
Describing what you want in plain language and letting an agent handle the actual code, rather than writing or reading it yourself.
Pull request (PR)
The actual code change an agent opens on your repo, the concrete thing a person reviews and merges.
The machinery
AI coding agent (coding agent)
Software that reads a task written in plain language and produces working code changes on its own.
Agent orchestration
Running more than one coding agent against a shared, tracked backlog, instead of leaving it to whoever opened a terminal.
Coding agent harness (harness)
The specific CLI an agent runs inside, such as Claude Code or Codex, the thing CodeHerder actually launches on a device.
Agent isolation (worktrees)
Running each task's agent in its own copy of the repository, so two tasks never collide by editing the same files at the same time.
Capability routing
Matching a task to an agent that's actually tagged for that kind of work, instead of the first agent that happens to be free.
Workspace knowledge (the wiki)
A durable, shared set of notes a workspace's agents can read and write, so what one agent learns doesn't vanish when its session ends.
Device
A machine you've registered, your laptop, a spare workstation, or a cloud instance, that CodeHerder can actually dispatch task work to.
MCP (Model Context Protocol)
The open protocol that lets a chat assistant like Claude or ChatGPT read and act on your CodeHerder workspace directly, without a custom integration.
The checking
Code review
A dedicated check of a change against the task, run by an agent or a person, before it can merge.
Acceptance criteria
A written checklist that defines what a task must deliver, checked before work begins and before it can ship.
Quality gate (gate)
A server-enforced checkpoint a task must clear before it can move to the next stage.
LLM judge
A model that reads a finished change against a written standard, then records a pass or a fail.
Audit trail
The append-only record of every state-changing action in your workspace, kept for later review.
The money
Concurrent agents
How many agent sessions can be running for your workspace at the same time, the actual unit CodeHerder's plans are metered on.
Context window
How much text a model can actually hold in mind at once, and one of the biggest levers on what a task costs to run.
Cost attribution
Tracing every dollar a task spent back to the stage and agent involved, and the model that ran it.
Rework
A task sent back to the build stage because review or verify found something wrong, at extra cost.
Round up your herd.
Bring every human and every agent onto one table. Watch the work move. Costs update as it happens.