Skip to content

Agents

Agents

Agents bring an AI assistant into the workspace. You author instructions and boundaries in source; the runtime and host do the rest.

Configure the agent

Two source files configure the agent: src/+agents.md — the system message of every agent turn — and an envoy at src/envoys/+<name>.ts

// src/+agents.md
What the collections mean, what the business does, and the house rules
for tone and escalation. This is the system message of every agent turn.

The prompt is shared by web chat and every envoy. An envoy adds its own task on top of it and names the policies it acts under — scope and capability come from those declarations, never from a per-agent data list.

Two doors

The same agent loop runs behind two doors:

  • web chat — the signed-in user acts on their own account
  • envoys — the envoy principal acts under its declared policies

Envoys

Reach the agent over Telegram or WhatsApp via envoys — each declares its transport, audience, the policies it acts under, and its own task.

Tools

The agent acts through tools, in three tiers:

  • Built-in tools — shipped with every bolt, always available
  • Custom tools — defineAgentTool declarations, one file per tool in src/capabilities/tools/
  • Host tools — trusted host-process operations a policy names, plus sandbox tools for in-session delegation (offered unless an envoy disables it)

Skills

Drop src/capabilities/skills/<name>/+skill.md files — one markdown file per directory — to give the agent durable domain knowledge; a skill reaches a turn only when a policy names it.