Agents

A Wacht agent is a reusable worker you configure once and reuse. You give it a model, tools, knowledge, and behavior rules. Then you point it at work:

• Chat with it — open a session and have a conversation.
• Give it a task — a single piece of work with a description (and optionally a file or two).
• Schedule it — recurring or one-shot work that runs without you in the loop.
• Share it — issue a session link so a teammate, customer, or another service can use the agent.

That's the whole product surface. The rest is how it's modeled.

What you'll actually do

• Model overrides — pick a different strong or weak model for one agent.
• Execution hooks — run setup or cleanup tools around every run.
• Approval policy — gate risky tool calls behind review.
• Scheduling — recurring or one-shot agent work.
• Composio integrations — wire Reddit, Gmail, Slack, and other third-party tools.
• Agent sessions — share a chat URL.

How it's modeled (skip unless you're driving the API directly)

The agent record doesn't carry tasks on itself. Work is modeled as:

actor                ← who the work runs as
└── actor project    ← binds an actor with an agent
    ├── board item   ← the actual task: description + optional schedule
    └── thread       ← the running conversation

This separation lets one agent be reused across many tasks and schedules. If you're calling the platform API directly (@wacht/backend, bench CLI), each layer is its own object — see the agent runtime reference.