you can't charm an agent

I now see a different cadence in our logs: a user retries the same task, fails, and disappears. It leaves no churn survey, support ticket, or angry tweet. The user is an agent, and in some products agents account for much of the traffic.

I shared this at a Mintlify event while discussing the agent experience work I lead at Auth0: context, tools, and evals. We publish a score at auth0.com/agent-experience because a team needs a number it can improve.

Developer tools have started serving this user through concrete surfaces. Cloudflare ships MCP servers and llms.txt files. Vercel makes docs queryable through llms.txt and exposes deployments through MCP. Sentry lets a model triage issues without opening the dashboard. Railway gives agents a capable CLI.

These products make information and actions available without depending on a visual interface.

AX inherits DX

Clean APIs, structured errors, stable SDKs, and precise docs serve agents because models learned from human code and use the same interfaces. Teams that invested in developer experience start with better agent experience.

A vague API pushes the model to guess. It may retry the wrong action or give the customer a false explanation. The human hears confidence while you find repeated failures in the logs.

I like the discipline this creates. An agent gives no credit for visual polish or brand trust. The logs reduce the experience to completion, retries, and recovery.

where to start

• Make the product headless. An agent needs a route that works without the UI.
• Publish machine-readable context. I added bharath.sh/llms.txt to this site. Keep it current with the docs.
• Expose actions. Choose the SDK, CLI, MCP server, toolkit, or skill surface that matches the product.
• Return structured errors. Tell the agent what failed, why, and which recovery action is safe.

Then add evals. Pick five tasks an agent should complete from a cold start and run them on each release. Evals catch a stale llms.txt, an SDK regression, or an MCP response that changed shape.

At Auth0, our open-source agent evals live at github.com/auth0/auth0-evals, and we publish their score at auth0.com/agent-experience. The scorecard also has an agent-readable version at auth0.com/agent-experience.md.