Runtime runs agents against real systems. A trigger arrives from Slack, email, a ticket, GitHub or a schedule, and the agent investigates across connected tools, does the work and reports what it found. Their own example: a duplicate charge queried in Stripe, checked against the database and the code, with a refund drafted for approval.
The controls are what make that acceptable in production. Database access is read-only with sampling and PII redaction, approval gates sit in front of sensitive actions such as refunds above a threshold, and every session shows each tool call and its cost. Agent memory recalls previous runs, so the second incident of a kind starts from the first.
Connectors cover Stripe, Zendesk, Datadog and GitHub, self-hosting is available, and it is SOC 2 with SSO. It starts free with no published tiers, aims at fast-growing technical companies rather than general teams, and an agent with read access to production data is a decision to make deliberately, however good the redaction is.
