Why this decision matters
Once an agent connects to business tools, it needs explicit access boundaries, action records, approval controls and failure fallback.
Tool-enabled agents can read enterprise data or trigger actions, so users, resources, actions, logs and high-risk approvals must be designed as part of delivery.
Conditions to confirm before development
- Least privilege by user, role, resource and action
- Logs for prompts, retrieval, model output and tool calls
- Human confirmation for high-risk or external actions
- Refusal and fallback on missing evidence, permissions or tools
Implementation and delivery approach
Treat the agent as a business-system component and design identity, tools, logs, approval and exception flows together.
Use normal, boundary, unauthorized, repeated and failed-action cases to verify controls and traceability.
Acceptance boundary
This guidance applies to the agreed data, system and environment. Project-specific scope, dependencies and acceptance conditions must be confirmed separately.