Sentry Seer

Debugging a production error has two halves, and only one of them is hard. Knowing that something threw is easy — a monitoring tool has told you within seconds for a decade.

The hard half is the reconstruction: what state was the application in, which recent change is implicated, why does this only affect some users, and what does the stack trace mean in the context of code you may not have written.

Seer's whole proposition is that the evidence for that reconstruction is already sitting in Sentry. This review covers why that data proximity matters more than model quality, what to make of automatically generated fix PRs, how the contributor-based charging behaves, and where a general coding agent like OpenCode does a different job.

Seer, In Five Lines

• Seer is an AI debugging agent built into Sentry that investigates errors, identifies root cause and opens pull requests with proposed fixes, rather than only reporting that something broke.
• Its advantage is data it already holds. Seer reasons over the errors, traces and logs Sentry collects, so it starts from evidence rather than from your description of the bug.
• It reviews pull requests too, checking a proposed change against the failure patterns your application has actually produced in production.
• It reaches you where you already are — Sentry's own UI, Slack alerts, and editors including Claude, Copilot and Cursor.
• Charging is a per-active-contributor add-on on paid Sentry plans, where an active contributor is anyone committing two or more pull requests to a connected repository.

Starting From Evidence Rather Than From A Description

This is the entire argument for Seer over a general-purpose coding assistant, and it is a strong one.

When you ask a general model to debug something, you describe the problem. You paste a stack trace, you explain what you expected, you supply whatever context you remember to include. The model reasons from your account of the situation, which is filtered through your existing assumptions about what went wrong — and those assumptions are frequently the reason you are stuck.

Seer skips that step. It reads the actual error, the actual stack trace, the actual trace spanning services, the logs around the moment of failure, and the breadcrumbs describing what the user did beforehand.

That is a meaningfully different starting position. The most valuable thing in a hard bug is usually a detail nobody thought to mention, and a system reading the raw telemetry has no reason to omit it.

It also explains why this product could only really be built by an error monitoring company. The model is a commodity; the ten years of instrumentation feeding it is not.

The natural-language querying against production data follows from the same foundation. Asking a question in plain English and getting an answer grounded in your own telemetry is more useful than the same question asked of a model with no access to it.

Ask a general model to debug and it reasons from your account of the problem — which is filtered by exactly the assumptions keeping you stuck.

Automated Fix Pull Requests, Assessed Fairly

Seer reads a stack trace and opens a pull request proposing a fix. That is the headline capability and it deserves a calibrated reading rather than either enthusiasm or dismissal.

Where it genuinely works is the large category of unglamorous bugs: a null check that was never added, an unhandled error case, a type mismatch, an off-by-one, a missing await. These are most production errors by volume, they are tedious rather than difficult, and they are exactly what a system with the stack trace in hand can propose correctly.

Where it will not work is anything where the fix depends on intent. If the error reveals that a business rule was implemented wrong, the correct change is a product decision, and no amount of telemetry contains it.

The risk to name is the specific one for automated fixes: a change that makes the error stop without addressing why it occurred. Wrapping a failing call in a try/catch removes the alert and preserves the bug, and a PR that does this looks superficially like a fix.

So the discipline is unchanged from any other patch. Read the diff, understand why the change works, and reject anything that only silences a symptom. A generated PR is a starting point that saves the reconstruction time, not a change you merge because a machine wrote it.

The pull request review feature is arguably the better half. Checking a proposed change against failure patterns your application has actually produced is something no reviewer can do from memory.

1. Turn it on for one project rather than the whole organisation.
2. Read the root cause analyses on issues you have already solved, and see whether it agrees.
3. Review the first ten generated PRs closely — you are calibrating trust, not saving time yet.
4. Reject anything that suppresses the error rather than fixing the cause, and note how often that happens.
5. Only then decide which projects should have it running by default.

Reaching You Where You Already Work

Seer runs inside Sentry's UI, in Slack alerts, and inside editors including Claude, Copilot and Cursor. That distribution decision matters more than it looks.

Debugging tools fail on adoption rather than capability. A tool that requires opening a separate dashboard, at the moment an engineer is already context-switching under pressure, gets used twice and then forgotten.

Appearing in the Slack alert means the analysis arrives with the notification — at exactly the moment attention is on the problem. Appearing in the editor means it is available where the fix will actually be written.

The editor integrations are also a quiet strategic statement. Rather than competing with Cursor and Copilot, Seer supplies them with production context they cannot otherwise reach. That is the right posture: an error monitoring company's advantage is the data, not the coding interface.

For teams already running an AI SRE platform, the distinction is worth drawing. Tools like Traversal reason about infrastructure dependencies and incidents across services; Seer is focused on the error and the code that threw it. They overlap at the edges and answer different questions.

Billed Per Active Contributor

This site does not publish plan prices. Seer's charging structure is unusual enough to be worth understanding, because it is not the seat model you would expect.

Seer is an add-on for paid Sentry plans — Team, Business or Enterprise — charged per active contributor per month. An active contributor is defined as anyone who commits two or more pull requests to a connected repository.

That definition is doing real work. It is not counting everyone with a Sentry login, which in most organisations is far more people than the ones writing code. Product managers, support staff and designers who look at error dashboards do not become billable.

The practical consequence is that cost tracks your engineering headcount rather than your monitoring footprint, and it flexes automatically as contributors come and go. For an organisation where many people watch errors and few people fix them, that is a favourable shape.

Two things to establish before enabling it: that you are on a qualifying paid plan, and how the two-PR threshold behaves across a month with contractors or occasional contributors. Ask specifically whether someone who commits twice in one month and never again is billed once or continuously.

The wider point for budgeting is that this is an add-on to a platform you are already paying for. The honest comparison is Seer's incremental cost against the engineering hours currently spent reconstructing what a stack trace means.

The Prerequisite Nobody States

Everything above assumes something that is worth making explicit: Seer is only useful if Sentry is already your error monitoring platform.

That is not a criticism, it is a scoping fact. If you run a different error platform, adopting Seer means migrating your instrumentation — which is a real project across every service you operate, and one that should be justified on its own merits rather than on a debugging feature.

There is a related dependency worth checking honestly: Seer is only as good as your instrumentation. An application with sparse error capture, missing traces or no breadcrumbs gives the agent little to reason over, and it will produce correspondingly thin analyses. Teams whose Sentry setup is a default install from three years ago should expect less than teams who have invested in it.

On data handling, Sentry states that AI-generated output is shown only to the requesting user, that customer data is not used for model training without explicit consent — disabled by default — and that processing runs through vetted subprocessors prohibited from using data for their own training.

That last set of statements matters more than usual here, because the data in question is production telemetry, which routinely contains more than teams realise. Confirm the terms with your own security review rather than the marketing page.

• Requires Sentry as your error platform — otherwise it is a migration decision, not a tool decision
• Analysis quality tracks instrumentation quality, so a thin setup gives thin answers
• Output is shown only to the requesting user, per Sentry
• Model training on customer data is off by default and requires explicit consent
• Production telemetry often contains more sensitive data than teams expect — review it

Capability, Coverage And Charging