AGENT SETUPHOW IT ALL WORKS
GitHub
…
Chapter 1

One Mac, seven agents, one rulebook.

This site explains how Isaac's AI coding agents are set up and how they work together. It is generated from the same files the agents read, so what you see here is what is actually configured, not a description that drifted.

Agents and hosts—runtimes plus the apps that launch them
Tool servers—in the shared catalog
Skills—loaded only when a task matches
Guard hooks—explained, with every runtime binding
Files—tracked and published in full

The setup in plain English

What lives here. Two AI coding agents do the work on this Mac: Claude Code (Anthropic) and Codex (OpenAI). OpenCode and Hermes stay installed but are outside the delegation graph, and Argent is a tool for testing phone apps. Each is a separate product with its own settings folder. This site and the repository behind it hold the shared rulebook and a copy of each product's settings, so the whole arrangement can be understood, audited, and rebuilt on another machine.

One rulebook, three layers. Every agent reads the same three files before it does anything: the constitution (rules that never change per task), the runtime overlay (which model orchestrates, which does routine work, which reviews), and the machine-readable routing graph. When prose and graph disagree, the graph wins.

Who is the boss. A Claude model (Fable 5.2 or Opus 5.5) or Codex orchestrates: it plans, splits the work, hands pieces out, checks results, and reports. Under Fable, Opus subagents implement; Codex takes testing and debugging, with Opus as its fallback; every subagent starts at medium effort. A review runs only when Isaac says yes, and it crosses vendors: work written by Claude is reviewed by Codex, and work written by Codex is reviewed by Claude, because a reviewer from the same family shares the author's blind spots.

When Isaac names a helper, that is final. "Use Luna" or "send this to Codex" means exactly that provider. An agent may not quietly substitute one of its own subagents because it was easier.

What is never done without asking. Committing, pushing, opening a pull request, deploying, publishing, running a database migration, posting to social media, writing to the marketing database, or spending money on a large media generation. Plain release language ("push it live", "ship it", "commit and push") is one authorization for the full chain; nothing weaker counts and nothing carries over from an earlier message.

Everything is shared. All agents see the same tool servers, the same skill library and authorization rules. A small lint runs at the start of each session and prints a line if any of that has drifted.

Experimental features are on. Every experimental feature in every agent is enabled, with a short, named list of exceptions. Codex's own in-app browser and Computer Use are deliberately on.

How it stays honest. Nothing here contains a password, token, cookie, or account. Every refresh runs two secret scanners and refuses to sync if either finds anything. A daily job refreshes the snapshot; committing and pushing remain a human decision.

Chapter 2

The agents

Seven runtimes with their own settings folders, and five apps that launch them and inherit everything. Each card says which models it runs, what its job is, how it reads the rulebook, and how its permissions are set.

Hosts that inherit the setup

Model roles, from the routing graph

Named agent definitions

Files under claude-code/agents and codex/agents. The two Haiku dispatchers (luna, codex) exist so a named external delegate is visible where the model looks for agents.

Chapter 3

The rulebook

Six files. Three are policy every agent loads; three are the executable graph, the tool catalog, and the skills policy they point at. Every one is readable here in full.

PRIMARY SOURCE OF TRUTH

Global agent constitution, every word

Source of truth. The complete ~/AGENTS.md as the agents read it. Rules appear exactly once; mechanics live in skills and load only when a task needs them.

Loading complete AGENTS.md…

Chapter 4

How work flows

In the order an orchestrator applies them. Every value below is read from the routing graph.

    Ceremony scales to blast radius

    The route a task takes

    Delegate Review

    How a named delegate is actually reached

      Chapter 5

      The words that trigger things

      What each phrase authorizes, how much ceremony it carries, and which file says so.

      Chapter 6

      Permissions and authorization

      Permission posture per runtime

      Authorization rules

        Requires Isaac's instruction

        Shared Git hook

          Every hook, with its runtime bindings

          Chapter 7

          Tools (MCP)

          One pinned catalog for every runtime. The default profile starts automatically; the others switch on for a task.

          Profiles

          Outside the shared catalog

          Chapter 8

          Browser automation

          The chain, from an agent's request to a tab in Isaac's Chrome

            The four transports

            Invariants

              Read the full design

              Chapter 9

              Experimental features

              Chapter 10

              Skills

              One canonical copy of each skill in ~/.agents/skills. Claude and Codex link to it. A skill's description is its trigger: every session sees the names and descriptions, and a skill's body loads only when a task matches (or Isaac names it). Isaac's private library of about 84 leaf skills is installed as seven category routers (ads, app-store, operations, seo, video, voice, website): a router fires on its subject and points the agent at exactly one leaf, read on demand, so the library costs about 1,000 tokens per session instead of 10,000.

              —
              Chapter 11

              Backup, refresh, drift

              How the live machine becomes this repository, how secrets are kept out, and how drift between the layers is caught.

                Security controls on this site

                  Recent snapshots

                  What the repository holds

                  Chapter 12

                  Every file

                  Every safe tracked Markdown file and registered agent descriptor, readable in full.

                  —
                  DOCUMENT

                  Loading document…

                  Download
                  Loading complete source…
                  Snapshot refreshed