How agent-ready is Looker?

Looker's agentability audit returned an overall score of 34 out of 100 on its homepage, with similar results across pricing (36) and documentation (24) pages. This places the product in the lower tier of agent-readiness, indicating that autonomous AI agents will encounter significant friction when attempting to parse information, complete workflows, or recover from errors on Looker's public web presence.

The audit identified three critical weak areas: clean handoffs (0/100), shadow UI avoidance (15/100), and control (20/100). Clean handoffs—the practice of providing structured, actionable error context—scored zero, meaning agents receive opaque failure messages with no programmatic path forward. Control mechanisms for pausing, canceling, or undoing actions are largely absent. Status reporting (30/100) and defaults (33/100) also fell below acceptable thresholds, while machine readability (64/100) and chunking (60/100) showed relative strength.

These gaps will prevent agents from reliably navigating multi-step workflows, recovering from invalid inputs, or managing long-running operations without human intervention. Addressing error handling, action reversibility, and progress transparency would materially improve agent operability.

Looker can improve agentability by focusing on the following concrete changes:

• Return structured error context in both UI and API responses, including what failed, an error code, and the offending field—replace generic 'Something went wrong' messages with actionable detail.
• Make every error message prescriptive by stating the next required action, for example: 'Email is invalid—enter a valid address' rather than only describing the problem.
• Gate destructive actions behind explicit, distinctly-labelled confirmation steps instead of same-styled buttons, and provide labeled pause or cancel controls (such as 'Cancel import') for long-running operations.
• Support undo or override functionality on completed actions, particularly for destructive changes, and allow agents to edit results rather than restarting workflows.
• Emit descriptive status text on asynchronous actions ('Uploading 3 of 10 files') and adopt enumerated multi-step or activity-feed patterns for long tasks instead of bare spinners.
• Surface confidence or certainty metadata on key outputs, such as a confidence field in API responses or 'verified vs best-guess' labels in the UI.