AI agents
Configure AI teammates with instructions, knowledge, tools, a budget and an approval queue.
An AI agent is a teammate you configure rather than hire. Where Molly works for whoever is asking, an agent has its own job: it's woken by events, works through tasks on its own, and asks a person before doing anything that needs a decision.
What an agent is made of#
| Part | What it's for |
|---|---|
| Instructions | The job, in plain words — Prepare renewal quotes, draft reminder messages and summarise cascade progress for the renewals team. |
| Model | Which model it runs on, for example GPT-4o |
| Knowledge | The slice of the knowledge base it may read — claim forms, policy wordings, SOPs |
| Tools | The actions it's allowed to take, such as looking up a policy or creating a task |
| Playbooks | Reusable SOPs and skill modules, assigned to one agent or several |
| Memories | Facts, preferences and instructions it keeps between tasks |
Agents can also connect to MCP servers for extra tools. Test the connection from the agent's settings to see which tools it exposes.
How agents work#
- Waking. An agent wakes when something happens: an event, a message, or a call to its inbound webhook. It triages each one, then acts on it or ignores it. The overview shows how often it woke in the last seven days and what it acted on.
- Tasks. Each piece of work is a task thread with a goal. A task can be one-off, a conversation, or standing — a recurring job on a schedule. You can pause, resume, cancel, edit or re-run a task, and message the agent inside it.
- Record pages. Ask agent on any record starts work about that record.
Approvals#
When an agent wants to do something that's gated, it creates an approval request and waits. Each agent's overview shows its pending approvals, and the queue offers four decisions:
- Approve — let it go ahead, optionally with a note;
- Edit — change what it proposed, then approve;
- Reject — stop it, with a note;
- Always allow — approve, and let this kind of action through for this agent from now on.
Budgets and usage#
Organisation-level agent settings include a monthly budget in US dollars, a per-task budget, the default model, whether agents may act autonomously, and whether agents are included in inbox auto-assignment. Each agent shows its cost this month against its budget, with the tokens behind it, broken down by day and by thread. Switch an agent off from the agents list at any time; its threads stay for review.
Through the API#
Agents live under /docs/api/agents. Wake one with an event:
POST /api/v1/agents/{id}/wake{ "source": { "kind": "renewal_due", "summary": "Policy MAXLIFE/2024/101608 renews in 30 days", "payload": { "record": "RECORD_ID" } }}Start a standing task that runs every weekday morning:
{ "goal": "Summarise yesterday's renewal replies for the team", "kind": "standing", "schedule": { "every": 1, "unit": "day", "at": "08:00", "enabled": true }}That body goes to POST /agents/{id}/tasks. The kind value in a wake source is free text, and the reference doesn't list accepted schedule units, so check /docs/api/agents before relying on them. Reading needs agents:view; configuring agents and deciding approvals need agents:manage.