Agents

Updated Aug 28, 2026View as Markdown

An agent is a model, a prompt, and a set of tools. You declare one with an [agent.<id>] section.

subs.toml

[llm.openrouter]
type = "openrouter"

[agent.oncall]
llm = "openrouter"
model = "anthropic/claude-sonnet-4-5"
system = "You are the on-call assistant."

The ID is how everything else names the agent. Slack routes to it. A client submits to it. A parent agent calls it.

Agent keys

Key	Meaning
`llm`	Which `[llm.<id>]` block the model call runs on.
`model`	The model to call.
`system`	The system prompt.
`effort`	How hard the model thinks.
`description`	What this agent does, for a parent that calls it.
`mcp`	Connections whose tools the model can call. See Connectors.
`plugins`	Plugins whose skills and servers this agent gets. See Plugins.
`subagents`	Other agents this one can call. See Subagents.
`tools`	Tools the browser runs. See Client-side tools.
`worker`	Where decisions go. See Workers.
`retry`	Timeouts and attempts. See Retries.
`signing_secret_env`	The variable holding the signing secret, for an engine you run.

Config has the full reference, including the defaults and the table forms of mcp and plugins.

Declare more than one agent

A second agent is a second section. Each one names its own model.

subs.toml

[llm.claude]
type = "anthropic"

[llm.cheap]
type = "openai"

[agent.assistant]
llm = "claude"
model = "claude-sonnet-4-5"
system = "Delegate research to the researcher, then answer."
subagents = ["agent.researcher"]

[agent.researcher]
description = "Finds and reads sources."
llm = "cheap"
model = "gpt-5-mini"
system = "Answer with citations. Be brief."

The model sees researcher as a tool. The child runs in its own session with its own transcript and cost.

Choose who decides

worker sets who answers this agent's decisions.

`worker`	Result
Set	The engine POSTs every decision to that URL.
Not set	The engine decides by accepting its own proposal.

Routing is per agent. An engine-hosted parent can call a worker-hosted child.

subs.toml

[agent.triage]
llm = "claude"
model = "claude-haiku-4-5"
worker = "https://triage.internal/agent"

Let a worker write the config

A section does two things: it declares that the agent exists, and it can set the agent's config. An agent whose worker builds its own config needs only the first.

subs.toml

[agent.support]
worker = "http://localhost:4000/substructure/agent"

That section sets no config, so session.start carries no proposal and the worker declares the whole agent.

Two rules follow:

An agent that sets no config needs a worker.
An agent that sets any config needs llm and model.

A partial config is a parse error.

Where tools come from

Source	Runs on	Declared in
A connector	The engine	`mcp` on the agent
A plugin	The engine	`plugins` on the agent
Your code	Your worker	The config the worker returns
The browser	The client	`tools` on the agent, with `handler = "client"`

Only browser tools go in the file. The tools that your worker runs are worker code.

Next steps

Connectors: tools from Sentry, GitHub, and any MCP server.
Plugins: skills and servers from a plugin directory.
Workers: decide with your own code.
Subagents: agents that call agents.
Config: every key in the file.