Tutorial 02 - Agents

An implicit agent like claude or codex is unconfigured: no custom prompt, adopting its provider's name, running the raw provider. This tutorial defines a custom agent with its own role and prompt, then slings work to it.

You should have my-city running with my-project rigged (from Tutorial 01).

Defining an agent

Each custom agent gets its own directory under agents/<name>/. Start by creating a rig-scoped reviewer:

~/my-city
$ gc agent add --name reviewer
Scaffolded agent 'reviewer'

~/my-city
$ cat > agents/reviewer/agent.toml << 'EOF'
dir = "my-project"
provider = "codex"
EOF

The agent.toml scopes the reviewer to my-project and switches it from the city's default claude provider to codex.

The provider catalog is explicit. gc init registered claude, but any other provider an agent references must also be registered in city.toml — otherwise every gc command fails with provider catalog is missing referenced providers. Register the builtin alias:

# city.toml — register the second provider
[providers.codex]
base = "builtin:codex"

(gc doctor --fix adds missing builtin provider aliases for you. base = "builtin:<name>" works for any provider Gas City ships a preset for. If you substituted claude above, skip this step — gc init already registered it.)

The agent needs a prompt. With no agent named, gc prime falls back to a generic worker prompt useful for a single-shot CLI invocation:

~/my-city
$ gc prime
# Gas City Agent

You are an agent in a Gas City workspace. Claim available work and execute it.

## Your tools

- `gc hook --claim --json` — find and atomically claim one work item
- `bd show <id>` — see details of a work item
- `bd close <id>` — mark work as done

## How to work

1. Claim work: `gc hook --claim --json`
2. Read the claimed bead and execute the work described in its title
3. When done, close it: `bd close <id>`
4. Check for more work. Repeat until the queue is empty.

gc prime shows the prompt an agent runs with — the instructions that tell it how to pick up and act on a slung bead. Pass an agent name to inspect a specific one: gc prime mayor prints the mayor's prompt; gc prime my-project/reviewer prints the reviewer's once we've written it.

The reviewer's prompt pairs the standard "find and execute" loop with review-specific instructions:

~/my-city
$ cat > agents/reviewer/prompt.template.md << 'EOF'
# Code Reviewer Agent
You are an agent in a Gas City workspace. Claim available work and execute it.

## Your tools
- `gc hook --claim --json` — find and atomically claim one work item
- `bd show <id>` — see details of a work item
- `bd close <id>` — mark work as done

## How to work
1. Claim work: `gc hook --claim --json`
2. Read the claimed bead and execute the work described in its title
3. When done, close it: `bd close <id>`
4. Check for more work. Repeat until the queue is empty.

## Reviewing Code
Read the code and provide feedback on bugs, security issues, and style.
EOF
$ gc prime my-project/reviewer
# Code Reviewer Agent
You are an agent in a Gas City workspace. Claim available work and execute it.
... # contents elided as identical to the above

You can also set the model and permission mode:

dir = "my-project"
provider = "codex"
option_defaults = { model = "o4-mini", permission_mode = "suggest" }

That file would live at agents/reviewer/agent.toml. Valid values come from each provider's options schema, so they differ per provider — o4-mini is a Codex model, while the same key on a claude agent would take values like sonnet or haiku.

An agent has five independent axes you can tune — the harness (provider), the model, who serves the model (upstream), the transport, and the runtime. For the full how-to, including running a harness against Bedrock, a proxy, or an OpenAI-compatible gateway, see Configuring an Agent.

Now that your agent is available, it's time to sling some work to it:

~/my-city
$ cd ~/my-project
~/my-project
$ gc sling my-project/reviewer "Review hello.py and write review.md with feedback"
Created mp-p956 — "Review hello.py and write review.md with feedback"
Auto-convoy mp-4wdl
Slung mp-p956 → my-project/reviewer

Because the reviewer is scoped to my-project, you target it as my-project/reviewer from inside that directory. Gas City started a Codex session, loaded agents/reviewer/prompt.template.md, and delivered the task.

There's no Attached workflow line this time: implicit agents like claude ship with mol-do-work as their default sling formula, but a custom agent has none until you configure one, so the bead is delivered directly. Sling still creates an auto-convoy to track it. Watch progress with bd show; when the work finishes, the review is on disk:

~/my-project
$ ls
hello.py  review.md

~/my-project
$ cat review.md
# Review
No findings.

`hello.py` is a single `print("Hello, World!")` statement and does not present a meaningful bug, security, or style issue in its current form.

Direct delivery suits fire-and-forget work. To watch an agent run or talk to one directly, you need a session — see the next tutorial.

What's next

You've defined an agent with a custom prompt, pointed it at a different provider, and slung work to it directly. From here:

• The six primitives — the canonical model agents, sessions, and work all build on
• Sessions — session lifecycle, sleep/wake, suspension, named sessions
• Formulas — how multi-step work should be done: steps, dependencies, and variables
• Beads — the unit of work; every task, message, and convoy member is a bead