Agent management
Enable agents, configure their gates, tune knowledge sources, and monitor their work.
The agent roster
Open Settings → Agents. Every ThorStack tenant comes with up to 25 specialist agents across 11 functions. Each agent has:
- A name and role (e.g. Forge — VP Sales).
- A scope — the modules and tools it can act on.
- A gate policy — what it can do unilaterally vs. what needs approval.
- A knowledge bundle — the documents and templates it has at hand.
- An on/off switch.
You only pay for agents you turn on.
Enabling an agent
Toggle the switch. The first run waits ~60 seconds for the agent runtime to schedule the agent's tick. From then on, the agent is "alive" — it'll act on tasks and events within its scope.
Configuring gates per agent
The default gates (see Approval gates) are fine for most teams. Per-agent overrides are useful when:
- A specific agent should never act unilaterally — set every gate to "always require approval".
- A trusted agent should bypass a normally-gated action — e.g. Ledger drafts journal entries without approval, but posting them still requires Finance lead.
Per-agent overrides are themselves audited.
Knowledge tuning
Each agent has a knowledge bundle — the corpus of notes, SOPs, brand voice, and reference docs it uses to make decisions. Bundles are scoped per-agent so agents see only what they should:
- Forge sees sales playbooks, deal templates, customer health rubric.
- Ledger sees chart of accounts, tax rules, vendor list.
- Echo sees brand voice guides, content calendar, audience definitions.
Update knowledge from the Notes module — checked-in changes propagate to the agent within a tick.
Memory layers
Each agent has four memory layers:
| Layer | Purpose |
|---|---|
| Working | The current task — flushed when the task ends. |
| Episodic | Recent runs — useful for "what did I tell you yesterday?" |
| Semantic | Long-term knowledge — names, accounts, SOPs. |
| Procedural | How-to memory — successful patterns the agent has seen work. |
You can inspect every layer per agent and prune what shouldn't be there.
Monitoring
The agent page shows:
- Tick health (last successful run, error rate).
- Cost over time (LLM tokens, tool invocations).
- Approval-gate volume (how often the agent asks for help).
- Top recent traces.
Anomalies (cost spikes, gate-rejection clusters) appear as alerts in your Notifications inbox.
Disabling
Toggling an agent off stops new ticks immediately. In-flight tasks complete to their next gate; gates that require this agent's approval re-route to the configured backup.
Next
- Approval gates — the gate model in detail.
- Bring your own LLM — provider routing per agent.
- RBAC — who can change agent settings.