Cole Medin demonstrates five agent controls in a 17-minute, MongoDB-sponsored tutorial using LangGraph and Atlas Agent Engine.

1. Enforce permissions at the tool boundary — 3:56

  • An allowlist determines which tools an agent can execute in a particular environment.
  • Medin presents two runs of a customer-deletion request: one follows the handbook; another attempts deletion and hits the platform’s policy block.
  • Remove unnecessary tools from the agent, then enforce permissions for tools that remain available through shared code or external servers.
  • A denied action can return a useful error so the agent can choose another path.

2. Pause consequential actions for approval — 6:42

  • The email example prepares a message, then waits for a person to review its body and recipients.
  • A reviewer can approve, reject or send feedback for another draft.
  • Medin suggests putting the approval queue in an existing workspace such as Slack, Teams or a custom dashboard.
  • Human review capacity limits which actions can require approval at scale.

3. Restrict outbound network access — 9:03

  • A sandbox isolates execution; a network allowlist restricts where the agent can send information.
  • Medin places deployment settings and permitted destinations in one YAML configuration.
  • A report to the fictitious Ridgeback Freight fails because its destination is blocked; the agent initially interprets that denial as a connection problem.

4. Keep memory current — 11:03

  • A demonstration recalls that Diego replaced Maya as a contact, despite a stale CRM record still naming Maya.
  • Memory needs updates and historical context, alongside keyword and semantic retrieval.
  • Medin distinguishes facts and preferences, past events, categories, and procedures as four memory types.
  • Editorial caution: retrieved memory also needs provenance and access controls; this example does not justify always overriding a live business record.

5. Trace each run — 13:58

  • OpenTelemetry-compatible traces expose model calls, tool inputs and results, and timings.
  • Medin uses the trace view to inspect handbook searches and discusses token counts and cost tracking.
  • These are creator-presented demonstrations, with no independent security or performance test.

Preview status matters

  • MongoDB’s documentation limits the preview to evaluation and prototyping, advises against production deployment, and anticipates breaking changes.
  • The production checklist can inform other systems; the demonstration platform is still a preview.

“Those rules are not guarantees.” — Cole Medin, 1:13