Platform / Agents
An agent is a job specification, not a chat window.
Each agent in the catalogue runs inside the platform's runtime — with defined inputs, stated guardrails, and a named human checkpoint for anything outside its scope.
Key concepts
What makes an agent production-ready.
Defined scope
Every agent has a stated job — not a general instruction to "help" with a domain.
Named guardrails
What the agent is explicitly not allowed to do is written down alongside what it does.
A human checkpoint
Every agent in the catalogue names who reviews its exceptions and approvals.
Catalogue
Six agents, written the way you'd describe a new hire.
This is a representative catalogue. Agents are configured against your workflows, not deployed generically.
Invoice triage
Reviews incoming invoices, extracts key fields, and routes exceptions.
- Inputs
- Invoices, purchase orders, vendor records.
- Guardrails
- Cannot approve payment or alter vendor records.
- Human checkpoint
- Finance review for exceptions and approvals.
12 processed today · 2 routed to finance
Support first-response
Drafts an initial response to incoming support tickets and flags ones that need escalation.
- Inputs
- Support tickets, knowledge base articles, account history.
- Guardrails
- Cannot issue refunds or make account changes.
- Human checkpoint
- Agent review before a response is sent for flagged tickets.
Awaiting reviewer approval on 3 drafts
Renewal watch
Monitors account signals ahead of renewal and prepares a summary for the account owner.
- Inputs
- Usage data, contract terms, support history.
- Guardrails
- Cannot contact the customer or change contract terms.
- Human checkpoint
- Account owner reviews the summary before any outreach.
4 accounts flagged for review this week
Vendor onboarding
Collects and validates vendor documentation against onboarding requirements.
- Inputs
- Vendor forms, tax documents, compliance checklists.
- Guardrails
- Cannot mark a vendor as approved or grant system access.
- Human checkpoint
- Procurement sign-off before a vendor record goes live.
1 vendor pending document resubmission
Ticket routing
Classifies and routes IT service tickets to the correct queue based on content and urgency.
- Inputs
- Ticket text, service catalogue, on-call schedule.
- Guardrails
- Cannot close tickets or change severity without confirmation.
- Human checkpoint
- On-call reviewer confirms routing for high-severity tickets.
38 routed today · 1 escalated
Report assembly
Pulls data from connected sources and assembles a draft operational report.
- Inputs
- Connected dashboards, spreadsheets, prior report templates.
- Guardrails
- Cannot publish or distribute a report without sign-off.
- Human checkpoint
- Report owner reviews and approves before distribution.
Weekly draft ready for review
Runtime
Where an agent's authority ends.
An agent's runtime enforces its stated guardrails at execution time — an agent cannot take an action outside its defined scope simply because a prompt asked it to.
Actions that fall inside a guardrail boundary but carry risk still route to the named human checkpoint before completing.
Define the first agent for your team.
A working session starts with one workflow, its inputs, and where a person needs to stay in the loop.