CRM for Agents is a relationship API—not a screen scraper

The dashboard is not the contract

Most CRM interfaces are optimized for a person who can interpret ambiguity. An autonomous client needs something stricter: typed operations, explicit preconditions, stable identifiers, bounded queries, and errors that distinguish authorization from validation and retryable failure.

Giving an agent a browser and hoping it behaves like a careful employee turns presentation markup into an accidental API. It is difficult to secure, hard to test, and nearly impossible to make reliably idempotent.

Make relationships legible to machines

A CRM for agents should model who acted, on whose behalf, for which workspace, and toward which actor. It should also distinguish an observation from a claim, a proposed change from an approved one, and an attempted action from a completed one.

• Actor identity and delegation chain on every request.
• Workspace-scoped reads with explicit field and record limits.
• Commands with idempotency keys and conflict-safe semantics.
• Policy evaluation before external side effects.
• Receipts that make consequences visible to people and agents.

Human usability remains the test

An agent-native CRM should not exile people. The same underlying record, policy, and audit event should support a calm human interface and a precise machine interface. That shared truth prevents an agent-only shadow system from growing beside the real customer history.