How to Have Coding Agents Pick Up New Sentry Issues Automatically (2026)
Start Claude Code or Codex on every new Sentry error so it reproduces the bug in a test and opens a draft pull request. Paradev automations, Sentry Seer, and GitHub Copilot compared.
For macOS on Apple Silicon and Intel. Linux AppImage is experimental.
Paradev was called Superset until October 2026.
To have a coding agent pick up new Sentry issues automatically, connect Sentry to a tool that starts an agent session when Sentry reports a new issue, and tell the agent to reproduce the error in a failing test before it changes any code. In Paradev, that tool is an Automation (Pro) with a Sentry → Issue created trigger. Each new issue starts Claude Code, Codex, or another CLI agent in a fresh workspace on a machine you choose. The agent gets the issue's title, link, and event payload, and leaves a draft pull request or a written note for a person to review.
Sentry also has its own agent, Seer. Its Autofix finds a root cause and can hand the fix to an external coding agent; Sentry's docs list Claude, Cursor Cloud Agents, and GitHub Copilot agents. Pick Seer when you want the analysis to happen inside Sentry. Pick a Paradev automation when you want your own agent and prompt to run against your own environment.
This guide is published by Paradev. Details about Sentry come from Sentry's docs as of October 2026.
What should the agent do with a new Sentry issue?
Most new errors with a clear stack trace are small bugs: a missing null check, a wrong type at an API boundary, an unhandled empty list. An agent can fix many of them. It can also write a confident fix for the wrong cause. A short rule keeps the output reviewable:
- Find the code in the stack trace.
- Reproduce first. Write a failing test that throws the same error. A test that fails before the fix and passes after it is the evidence a reviewer needs.
- Stop if it cannot reproduce. Write what it checked to a file, and do not guess at a fix.
- Open a draft pull request, linked to the Sentry issue, so a person checks it before anyone requests review or merges it. CI can still run on draft pull requests.
How do I set it up in Paradev?
- Connect Sentry under Integrations, and install and connect the Sentry plugin so the agent can read the issue's events and stack trace. Plugin tools reach Claude Code and Codex; other agents need their own Sentry MCP setup.
- Go to Automations → New automation. Pick the project for the repository that the Sentry project reports from, a device that stays online, and an agent.
- Under Triggers, click Add Trigger → Sentry → Issue created. Pick the Sentry project and set the level to Error and Fatal, so warnings do not start runs. Press Save.
- Write the prompt:
- When a run finishes, open it from Run History and read
TRIAGE.mdor the draft pull request. Mark the PR ready for review when you agree with it.
From the CLI, list the Sentry projects the trigger can watch, then create the automation:
Without --host, the automation runs on the machine where you type the command. Pass the ID of a host that stays online, from superset hosts list.
The trigger's event payload describes the issue, but the stack trace, breadcrumbs, and other events for the same issue come from Sentry itself. The Sentry plugin connects the agent to Sentry's MCP server for those, and Paradev keeps the Sentry credentials, so the agent and the machine never see them.
Which Sentry events should start a run?
| Trigger | Use it for |
|---|---|
| Issue created, level Error and Fatal | Every new error, the default setup above |
| Issue unresolved | Regressions: a bug that came back after it was marked resolved |
| Issue assigned | Only the issues someone hands over, for teams that want a person to choose |
| Any issue event | Rarely: it fires on resolve and archive too, so it needs a prompt that checks the event type |
Start narrow. A noisy Sentry project can create dozens of issues an hour, and each one is an agent run on your machine and your model subscription. Issue assigned is a good first step: a person triages in Sentry and assigns the issues worth a run.
What should I know before I leave it running?
- The device must be online. Runs fire on the device you picked. If it is asleep or offline when Sentry reports the issue, the run is recorded as missed and is not retried. Use a machine that stays on, such as a Mac mini or a Linux server running
superset start --daemon, or pick Cloud as the device if your account has cloud workspaces. - Runs can be delivered twice. Delivery is at least once. The "if a pull request already exists, stop" line catches most repeats, but two runs that start at the same moment can both pass it, so a duplicate pull request is still possible. Close the extra one.
- A created run is not a fixed bug. The run history shows that the workspace was created, not whether the agent's fix is right. Read the test and the diff.
- Production data. The event payload can include request data or user details that Sentry captured. It goes into the agent's prompt, so set Sentry's data scrubbing to match what you are willing to send to your model provider.
How does this compare with Sentry Seer and other options?
| Option | Who does the analysis | Who writes the fix | Where it runs |
|---|---|---|---|
| Paradev automation on Sentry issues | Your agent (Claude Code, Codex, or another CLI agent), with your prompt | The same agent | Your machine, a shared host, your own Linux server (Bring Your Own Cloud), or a cloud workspace |
| Sentry Seer Autofix | Seer, with Sentry's traces, logs, and profiles | Seer, or a handoff to Claude, Cursor Cloud Agents, or GitHub Copilot agents | Sentry and the chosen agent's environment |
| Sentry MCP server in your own agent | Your agent, when you ask it | Your agent | Wherever you run it, started by hand |
| GitHub Copilot cloud agent | Copilot, in a session started from a GitHub issue | Copilot | GitHub |
Seer's strength is context: it reads the traces, logs, and profiles Sentry already has, and adds its root cause analysis to the issue page. Pull request creation can be turned off in Seer's settings if you want a person to write the fix. A Paradev automation's strength is control over the run: you choose the agent and the prompt, the run happens in a full checkout of your repository with your tools and test setup, once the host or cloud environment has them, and the result lands in a workspace you can open, continue, or hand to a teammate. The two also combine: let Seer find the root cause in Sentry, and use a Paradev Issue assigned trigger to start your agent on the issues a person hands over.
Frequently Asked Questions
Can Claude Code fix Sentry errors automatically?
Yes, with something to start it. A Paradev automation with a Sentry Issue created trigger runs Claude Code on each new issue. Sentry Seer can also hand an issue to a Claude agent from its Autofix flow.
Will the agent push a fix to production?
No, unless your prompt and permissions allow it. The prompt in this guide opens a draft pull request, and your normal review, CI, and deploy steps still apply.
Does it work with Codex instead of Claude Code?
Yes. Pick Codex as the automation's agent, or pass --agent codex on the CLI. The trigger and prompt stay the same.
What happens when Sentry reports the same error many times?
Sentry groups repeated events into one issue, so Issue created fires once per issue, not per event. A duplicate delivery can still happen. The "already exists" check in the prompt stops most of them, but not two runs that start together.
Which plan do I need?
Automations are a Pro feature, and every trigger used in this guide works on Pro. See pricing.