Cloudflare opens Workers Issues in beta, and it can send a grouped error with stack trace and logs straight to Claude Code, Cursor or Devin

The gap between a stack trace appearing in production and a pull request that fixes it has been a person copying context into a prompt. Cloudflare said on 30 September that its Workers platform now has a built-in error monitor called Issues, in open beta, that groups a Worker's repeated failures into one issue and can hand the whole issue (error, stack trace, logs, traces, Worker version) to a coding agent through a configured automation.

The setup is one line of configuration in cloudflare.config.ts. There is no SDK to install, because Issues is built into the Workers runtime. Once on, it records uncaught exceptions, failed invocations, HTTP 5xx responses, output from console.log() and console.error(), and logs that carry a stack trace. It also flags runaway alarm conditions and code that writes large volumes of logs inside loops. Every failed request has its own request ID, but Issues groups them by underlying bug and shows when the error first appeared, how many times it has happened, and whether it is growing.

The handoff to the agent

When an issue crosses an occurrence threshold or comes back after a quiet period, an Automation sends it out. Cloudflare lists three built-in destinations for a coding agent, each named:

• Claude Code, connected with a routine ID and token.
• Cursor, connected with an automation webhook URL.
• Devin, connected with an API token and organization ID.

Anything else goes through a generic webhook or a chat and incident-management route. What the automation sends is the failure summary and diagnostic context captured with the issue: the exception, error, source-mapped stack trace, leading and trailing logs and traces, Worker version, and any application context added through the Workers runtime's built-in OpenTelemetry API. For deeper investigation the post says to connect the agent to Cloudflare MCP separately so it can query the related logs and traces, then propose a change and open a pull request.

Deploy control stays with the person. Cloudflare's framing is that the agent proposes; the review, deploy and mark-resolved steps are still in the human's hands.

The proof point Cloudflare offers

Cloudflare Workflows, the platform's own primitive for long-running multi-step applications, was the internal test. The team says that within a day of turning Issues on, it found two problems hidden inside a large volume of traffic:

• A control-plane migration stuck in a retry loop against a SQLite foreign key error in an edge case.
• A deletion process that exceeded a Workers subrequest limit under specific conditions and never finished.

Both are real bugs surfaced by the tool and both were fixed. The post does not report a broader before-and-after comparison, so treat that as a demonstration rather than a measured claim.

What is different about this shape

Coding agents have been able to read observability data, change code, run tests and open pull requests for a while. The missing step was the structured handoff: recognising that repeated failures come from the same bug, gathering the relevant logs and traces, sending the right context to the agent, and checking whether the fix worked. Without that, the agent must reconstruct the scope of the failure from raw telemetry before it can investigate.

Issues is Cloudflare replacing the person in that middle step, on its own platform, for Workers. The size of the effect depends on how much of a team's application runs on Workers and whether the agent's proposed changes hold up in review. The beta is open, and the configuration is the one line above plus the coding-agent credential for whichever tool the team already uses.

Source

• Detect and send production issues straight to your agent, Cloudflare blog, 30 September 2026.