Agents

An Enverif agent is a durable configuration, not a one-off prompt. It combines instructions, model selection, skills, connectors, execution bounds and capability policy.

Instructions

Define the objective, evidence requirements, disqualification criteria, output format and escalation behavior. Keep secrets out of instructions; credentials belong in encrypted model or connector connections.

Bounds

• Maximum turns prevents unbounded loops.
• Maximum runtime stops runs that exceed the configured wall-clock window.
• Maximum estimated cost uses the optional per-million-token prices on the selected model connection. 0 means no cost ceiling.

Capability policy

Risk categories are read, internal_write, network, external_write, secrets and destructive. Explicit deny patterns outrank all other choices. Secret access always asks. Destructive operations are denied by default and still ask when destructive capability is enabled.

Durable runs

Runs persist messages, tool calls, results, token counts, cost estimates and approval state in MySQL. Redis queues schedule continuation work; Redis is never the source of truth for run history.