Ghostbots.art

Guide 00 · Start here

The basics

The terms the other guides and the exhibits use. Each one in plain words, how we set it up, and where to read more.

Related:
The settings behind our agents
Checked against the source:
Nov 4, 2026
Tools:
OMP 18.14.1 · Claude Code 2.1.281 · Git 2.52

What it is for

An AI coding agent is more than a model in a chat window. Around the model sits a program that gives it tools, loads its instructions and decides what it may do without asking. The other guides assume you know these parts. Our agents run in OMP; where Claude Code works differently, we say so.

The terms at a glance

What it is

Model and role · Agent and harness · Tools and approval

What it knows

Session and context · Context files · Memory · Skills

How work is split

Subagents · Worktrees

What runs on its own

Hooks

One session, step by step

  1. Start

    You start the harness in a project folder. A new session begins with an empty context window.

  2. Load

    It loads the context files, the rules, and the name and description of each skill. With memory on, matching memories are added to the first turn.

  3. Ask

    You give it a task.

  4. Work in a loop

    1. The model picks a tool
    2. Checks: a hook can block the call, and the approval mode decides whether you are asked first
    3. The tool runs
    4. A hook can see the result
    5. The result goes back to the model

    This repeats until the task is done or you stop it.

  5. Fill up

    When the context window gets full, the harness compacts it: the older part becomes a summary.

  6. End

    The session is saved as a file you can resume. With memory on, parts of it can come back in later sessions.

Simplified, from the OMP and Claude Code documentation, Nov 2026. Each link goes to the term below.

What it is

Model and role

In plain wordsThe model is the part that reads and writes text: it reasons about the task and decides the next step. On its own it cannot open a file or run a command. A role is a name for one kind of work, such as quick searches, planning or review. Each role points to a model, so you change the model in one place.

In our setupOur OMP config has 24 roles. The commit review has its own role, on a different model from the one that writes the code. We do not name the models on this site.

More: Models and roles in Settings · Guide 01, Commit review by a second model

Agent and harness

In plain wordsA harness is the program around the model. It gives the model tools, loads its instructions, keeps the conversation and asks you for approval when its settings say so. An agent is the model at work inside a harness: it gathers context, acts, checks the result and goes again until the task is done.

In our setupOur agents run in OMP. We also use Claude Code, Codex and Antigravity, and all four load one set of rules.

More: How your agent behaves in Settings · Guide 03, One set of rules for four agent tools

Tools and approval

In plain wordsTools are the actions a harness offers the model: read a file, search, edit, run a shell command, start a helper agent. The approval mode decides which tools run on their own and which ask you first. Rules can make single commands stricter. Claude Code calls this a permission mode.

In our setupOMP runs in yolo mode, where nothing asks. Deny rules and a hook block recursive deletes instead. That is not a sandbox: an allowed command has the same file and network access as the person who started the agent.

More: Permissions · Sandbox and network in Settings

What it knows

Session and context

In plain wordsA session is one conversation with the agent. It is saved as a file, so you can resume it. The context is everything the model sees in that session: your messages, the files it read, command output and its instructions. The context window is how much of that fits. Each new session starts empty. When the window fills up, the harness compacts it into a summary, and details from early in the conversation can get lost.

In our setupOMP compacts at 50% of the window. Before that, at 30% and 40%, an extension tells the agent once to finish its step and suggest a handoff to a fresh session.

More: Context and instructions in Settings

Context files

In plain wordsFiles the harness loads into every session on its own, such as AGENTS.md and CLAUDE.md. They hold the standing rules: how to work in this repo, which commands to run, what never to do. One sits at user level and applies to all your projects; each repo can add its own. OMP reads both file names. Claude Code reads CLAUDE.md; its documentation says that from version 2.1.277 it also reads a repo's AGENTS.md when there is no CLAUDE.md (checked Nov 4, 2026).

In our setupOne source file holds our working agreement. OMP and Claude Code import it with a one-line @ reference; Codex and Antigravity get a copy. In each repo, AGENTS.md holds the rules and CLAUDE.md only imports it.

More: Guide 03, One set of rules for four agent tools · Context and instructions in Settings

Memory

In plain wordsNotes the agent keeps between sessions, such as decisions and habits, so they can come back later. Memory is background, not an order: when it disagrees with the repo or with your message, those win.

In our setupOMP has memory off by default. We turned it on, with the memories stored in a database on our own machine. The commit review runs with memory off, so it judges the diff and not old notes.

More: 04 · Memory across sessions (exhibit) · Memory in Settings

Skills

In plain wordsA skill is a folder with a SKILL.md file: instructions for one kind of work. At the start, the agent sees only each skill's name and description. It reads the whole skill when a task calls for it. A skill fits a procedure you would otherwise paste into the chat again and again.

In our setupOMP loads our own skills and also the ones in Claude Code's user folder.

More: Guide 04, Skills: instructions an agent loads when it needs them · Skills in Settings

How work is split

Subagents

In plain wordsA helper agent that the main agent starts for one part of the work, such as searching the code or reviewing a change. It has its own context and returns a short result.

In our setupAt most four run at once. In OMP, subagents do not get the AGENTS.md files; rules that every helper must follow go into a separate rules file, RULES.md.

More: 06 · Parallel subagents (exhibit) · Subagents in Settings

Worktrees

In plain wordsA git worktree is another working folder of the same repository, with its own branch and its own files. Two agents can work on two tasks at once without editing the same files. Git lets one branch be checked out in only one worktree at a time.

real commands · folder and branch names are illustrative
git worktree add -b task-a ../repo-task-a   # a new folder on a new branch
git worktree list                           # both folders, each on its branch
git worktree add ../again task-a            # refused: task-a is already checked out

In our setupEach task gets its own worktree, created with Orca, and the agent sessions for that task run in it. OMP can also give a subagent its own copy of the repo.

More: Guide 02, Plan, review, implement · 05 · One directs, three work (exhibit)

What runs on its own

Hooks

In plain wordsA hook is code that runs on its own at a set moment: before a tool runs, after it, when a session starts or ends, or before a git commit. A hook can stop the action; a rule in a context file cannot. Harness hooks see the agent's tool calls. Git hooks see git itself, so they work with any agent.

In our setupOur OMP hooks block recursive deletes and add a log line to each commit. A git pre-commit hook in our knowledge repo hands each commit to a second model for review.

More: Guide 05, Hooks: which one, for what · Guide 01, Commit review by a second model · 03 · A hook that keeps a log (exhibit) · Hooks in Settings

What to watch out for

Where to go next

If you are new, read the guides in this order. It goes from what the agent reads, to what runs on its own, to how several agents share the work.

  1. One set of rules for four agent tools: the context files, set up once for every tool.
  2. Skills: instructions an agent loads when it needs them.
  3. Hooks: which one, for what. Code that runs on its own and can stop an action.
  4. Commit review by a second model: a git hook that lets another agent stop a commit.
  5. Plan, review, implement: three sessions per task in one worktree.
  6. Overnight batch, one morning overview: several agents work on separate branches overnight, and in the morning you decide from one overview.

Sources

← All guides