A Cortex agent is an AI worker defined by a playbook: a role, a list of steps, an allowlist of tools, the knowledge it may consult, and rules for when it must hand off to a person. Agents run inside Limnn, so they act on the same records as your team in Sales, Support and Customer Success, under the same workspace boundary. A swarm is a group of specialist agents coordinated by a supervisor.
The anatomy of a playbook
- Engine: SUPPORT, SALES or CS. This decides which tool catalog the agent can draw from.
- Channels: CHAT, VOICE and/or EMAIL. The same playbook can serve several channels.
- Steps: the ordered procedure the agent follows, written in plain language.
- Tool allowlist and per-tool config: only allowlisted tools exist for the agent. Config passes limits such as a refund cap (for example capUsd: 50).
- Knowledge sources and web search: which knowledge the agent may ground answers in, and whether it may search the public web.
- Model: auto, sonnet or opus. Auto lets Cortex pick based on the task.
- Tone, SLA seconds, confidence-to-send and spend cap (in credits).
- Escalate rules: conditions that force a hand-off to a human.
Tools and risk levels
Every tool is tagged READ, WRITE or RISK. READ tools (search_knowledge, lookup_ticket, ticket_history, lookup_deal, account_history, check_suppression, web_search) run automatically inside the agent loop. WRITE and RISK tools (reply_to_ticket, set_ticket_status, add_internal_note, log_sales_activity, advance_deal_stage, draft_outreach, escalate_to_human) are staged as proposals with a one-line summary instead of executing immediately.
Revenue and data-quality agents add their own tools on the same model: find_overdue_invoices, find_duplicate_records and find_incomplete_records are READ; draft_invoice_followup and propose_merge are RISK; propose_field_fix is WRITE. A follow-up draft is queued as a scheduled collections step only after a person approves it, and it is never sent by the agent itself.
Autonomy levels
- DRAFT: every WRITE or RISK action is staged for human approval. Start here.
- AUTO_LOW: low-risk WRITE actions can execute automatically; anything riskier is still staged.
- FULL: the agent may execute permitted actions itself. Use only for well-tested playbooks with a spend cap.
Set up and run an agent
- 1Open Cortex from the main navigation. /cortex is the Agent Hub (see the Cortex Agent Hub article).
- 2Go to Templates and install a ready-made agent, or go to Agents and choose Build your own. Installed templates start switched off until you review them.
- 3Open the agent's playbook, choose the engine and channels, and write the steps.
- 4Allowlist only the tools the agent genuinely needs.
- 5Attach knowledge sources (and fill in the Context tab) and decide whether web search is allowed.
- 6Set autonomy to DRAFT, set a confidence-to-send threshold and a spend cap, and add escalate rules.
- 7Switch the agent on, let it run against real tickets or deals, and review the staged proposals in the Inbox.
- 8Approve, edit or reject each proposal. Your decisions feed back into the agent as feedback.
Swarms and supervisors
In a swarm, a supervisor triages the incoming work and delegates to specialist agents, for example a Support resolver, a Sales follow-up agent and a Customer Success check-in agent. Each specialist keeps its own tool allowlist, so delegating never widens what any one agent can do. Voice agents use the same playbook model for spoken conversations.
Every run is recorded
A run ends in one of four states: AWAITING_APPROVAL, RESOLVED, ESCALATED or FAILED. It records the reply text, confidence, whether the answer was grounded in sources, cited sources, the number of steps, the model used and the credit cost.