The AI workforce
How the eight specialist agents operate, tick cadence, the 47-tool surface, scoped permissions, memory layers, autonomous reviews, and gates.
What the AI workforce is
ThorStack ships with eight specialist agents across eight functions, one lead per function rather than a sprawling org. Each agent has a name, a role, a scope, a knowledge bundle, and an on/off switch. You only pay for the agents you turn on. This page covers how they actually run; to configure them, see Agent management.
The roster
| Function | Lead | What they own |
|---|---|---|
| Executive | Atlas (CEO) | Strategy, prioritization, the weekly executive brief. |
| Sales | Forge (VP Sales) | Pipeline hygiene, deal coaching, follow-ups. |
| Marketing | Echo (CMO) | Content calendar, brand voice, channel ops. |
| Finance | Ledger (CFO) | Books, invoices, cash flow, tax filings. |
| Legal | Counsel (GC) | Contracts, NDAs, compliance posture. |
| Operations | Aria (COO) | Fulfillment, inventory, channel reconciliation. |
| People | Pulse (CHRO) | Onboarding, culture, performance cadence. |
| Data & Security | Cipher (Security) | Access reviews, audit logs, anomaly detection. |
How an agent runs
Each enabled agent runs on a 60-second tick. On every tick it checks its queue, tasks assigned to it, events in its scope, scheduled reviews, and acts on what it finds. Between ticks it's idle and costs nothing.
Work reaches an agent three ways:
- You ask, in chat or by assigning a task.
- An event fires, a deal moves, an invoice goes overdue, an order ships. See Event triggers.
- A schedule comes due, a daily or weekly review the agent owns.
Tools
Agents act through a shared surface of 47 built-in tools: 22 universal tools every agent can use (search, read a record, draft a message, write a note, create a task) and 25 native app tools scoped to specific modules (send an invoice, score a lead, schedule a post, send for signature). An agent can only call the tools inside its scope, and every call is written to the trace and the audit log.
Scoped permissions
An agent never sees more than it should. Its scope is the set of modules it owns, enforced the same way human permissions are, at every read, write, and tool call (see RBAC). Forge can't read the ledger; Ledger can't post to social. Visibility is checked server-side, including for the agent's own retrieval.
Memory
Each agent carries four memory layers, so it improves with use instead of starting cold every time.
| Layer | Holds |
|---|---|
| Working | The current task; cleared when the task ends. |
| Short-term | The current session and recent runs. |
| Long-term | Accumulated knowledge across the tenant: names, accounts, decisions. |
| Learning | Outcomes that feed back to improve future runs. |
Memory is per-tenant and never trains a shared model. You can inspect and prune any layer from Agent management.
Autonomous reviews
Beyond reacting to you, agents run scheduled reviews their function owns, for example a morning portfolio pass and a weekly briefing. These produce drafts and surfaced risks, not silent actions: anything with a financial, legal, or reputational blast radius still routes through a gate.
Approval gates
Sensitive work, outbound mail, payments, contract changes, bulk deletion, waits for a human. The agent prepares the action and holds it at a gate with the full context; you approve, edit, or reject. Defaults are conservative and tunable per agent. See Approval gates.
Agents and workflows
Chat and tasks are how you reach an agent directly. Workflows are how you wire agents into repeatable, multi-step automations with branches and gates. The same agents, tools, and gates power both; a workflow is just a pre-drawn path through them.
Next
- Mission Control Chat, talk to the workforce in plain language.
- Agent management, enable agents, tune gates and knowledge.
- Approval gates, the human-in-the-loop model.
- Workflow basics, chain agents into automations.