Tool Call Policies

Control which external tools and functions your agents can invoke, require approval for the dangerous ones, and keep agents on-mission.

Features

Tool Whitelists

Define exactly which tools each agent can call. Block access to sensitive operations like database writes or file system access.

Per-Tool Approval

Require human approval for only the sensitive tools (exact name or "prefix*" wildcard, e.g. delete_*) while everything else auto-runs — instead of an all-or-nothing approval policy. Matching calls pause for operator review; the rest proceed. Matching is case-insensitive so a case-variant tool name can't dodge the gate.

Intent-Drift Enforcement

Every tool call is compared (embedding cosine similarity) against the agent's declared purpose. In shadow mode drift is logged; in enforce mode a call that wanders off-mission is blocked (403). Fails closed if enforcement is on but no purpose is set, and audits every case it can't evaluate — no silent bypass.

Approval Workflows

High-risk tool calls pause for human review with 4-eyes (reviewer ≠ requester), a per-agent quota, and TTL expiry. Reviewers act from a live dashboard queue.

Hard Kill Switch

Stop a runaway agent instantly — per-agent or tenant-wide. A hard kill also revokes the agent's keys, so a raw gateway call on a still-valid key is rejected on its next request, not just the active sessions.