Sentry for AI agents: what it means

the short answer

"Sentry for AI agents" means bringing the error-tracking loop to failures that do not throw exceptions: capture the full session, detect bad behavior, group related failures, alert an owner and track the fix. Sentry covers agent traces, tool errors, latency and cost. Behavioral failures such as an incorrect result, skipped step or unsafe action need evaluations or failure analysis on top of those traces.

Origin of the phrase: Raindrop's YC W24 launch: "Sentry for AI Products"
What carries over: Capture with context, grouping into issues, alerts, triage
What breaks: No exception to catch, no stack trace to group by
What fills the gap: Evaluations and failure analysis that detect and group behavioral problems

Where the Phrase Comes From

Raindrop (raindrop.ai) launched at YC as "Sentry for AI Products": "AI engineers use Raindrop to get alerts about hidden issues and wins in their AI products." The argument was that conventional software reports its own failures through exceptions, and AI products fail silently - its launch named "Assistant Forgetting, Laziness, Task Failure, User Frustration" as the failure modes, none of which raises anything. Raindrop has since moved on to agents, but the phrase stuck, and it is a good description of what a whole category is trying to build.

What Carries Over from Error Tracking

The error-tracking loop is worth copying almost entirely. Its parts map onto agents cleanly:

Capture with context. A crash report carries a stack trace and breadcrumbs. For an agent, the equivalent is the session trace: the prompt, each tool call and result, the model outputs, in order.
Group occurrences into issues. Nobody reads ten thousand events. Grouping many failing sessions into one issue with a count is what makes the data readable.
Alert on new and spiking issues. A new failure type or a jump in an existing one should reach someone, once, with a link.
Triage as a workflow. Assign, discuss, resolve, and notice when a resolved issue comes back after a deploy.

What Does Not Carry Over

Everything above assumes a failure announces itself. Agents break that assumption in four ways.

There is no exception to catch. The agent told the user the refund was processed when it was not. Every call succeeded. Known failures can be detected with code checks, classifiers or LLM judges. Unknown failures require analysis across production sessions.
There is no stack trace to group by. Error trackers fingerprint where the code broke. Two agent failures can mean the same thing while sharing no text, tool sequence or exception, so grouping has to use behavior and context.
The fix is often not a line of code. It is a prompt, a tool description, a missing guard, or a model version. "Go to the line that threw" does not exist.
Detection costs money per event. Catching an exception is free. Running a judge over every session costs tokens, so teams sample, and sampling means some failures are never seen.

The fixing loop also changes. A finding may lead to a prompt update, tool redesign, evaluator, model rollback or behavioral policy. High-risk actions sometimes need prevention in the runtime, but most failures first need enough evidence to identify the right fix.

What Sentry Itself Now Covers

Sentry has built agent support into the product. Its agent tracing records "every LLM call, tool execution, MCP interaction, and handoff as a span with its inputs, outputs, token counts, cost, and timing", and auto-instruments the OpenAI Agents SDK, Vercel AI SDK, LangChain, LangGraph, Pydantic AI, the Anthropic SDK and Google GenAI, across Python, Node.js, browser, .NET, Ruby and other platforms.

Sentry can show that a tool errored, a provider timed out or a run became slow and expensive. The harder agent failures complete without an exception: the task was left unfinished, the wrong tool was used, a policy was ignored or the user was given a confident but incorrect result. Catching those failures requires evaluations or failure analysis that judges the agent's behavior and outcome, not only the health of the underlying calls.

Which Tools Fill the Gap

Tool	Behavior detection	Grouping and investigation	Alerts	Fixing workflow
Sentry	Operational errors, latency and cost; behavioral checks require additional instrumentation	Exception-based issues and trace investigation	Existing Sentry alert workflow	Code or configuration change
Raindrop	Plain-language Signals	Stumbles, ranked Issues and Triage Agent analysis	Slack	Investigate and compare a changed cohort with Experiments
Latitude	Conditions, LLM judges and JavaScript evaluations	Traces, evaluations and Signals	Email and Slack monitors	Agent Dispatch can open a fix PR
Judgment Labs	Agent Judge, Code Judge and Behavior Discovery	Cited judge results and behavior monitoring	Alerts, webhooks and Slack	Update rubrics, datasets or application code
Failproof AI	Code-based and LLM-based evaluations plus automated failure analysis	Related evidence grouped into findings with recommendations	Email, Slack, webhooks and dashboard	Assign the finding, validate a fix and optionally enforce a policy

Based on each vendor's published product information.

Raindrop stays close to the original “Sentry for AI” idea with Signals, recurring Issues and a Triage Agent. Latitude connects traces and evaluations to a coding agent that can open a fix PR. Judgment Labs focuses on judge quality and cited evidence. Failproof AI combines evaluations with automated failure analysis, findings, ownership and alerts, then lets a validated fix become a behavioral policy when prevention is required.

Known failures still benefit from explicit checks. A code-based evaluation can catch tool errors, missing outputs or exceeded limits without a model call. An LLM judge can score behavior that needs interpretation. Automated audits add the other half of the workflow by finding repeated patterns across sessions, including problems the team did not already encode as an evaluator.

Each result remains connected to the session and tool calls behind it. That matters during triage: an owner can move from the finding to the evidence, test a proposed change against real failures and monitor whether the same behavior returns after deployment.

When the Sentry You Have Is Enough

If you already run Sentry and your agent is young, start there. Turn on agent tracing, then turn the checkable silent failures into real errors: after each run, check the outcome in code - did the tests run, did any tool result carry an error, did the agent produce the file it was asked for - and report a failed check as an error event. You get grouping, alerting and assignment from a tool your team already watches, for no new vendor.

You outgrow that when the failures that matter stop being checkable in code - tone, correctness, a task that was technically finished and practically wrong - or when one of them is expensive enough that you want it stopped rather than reported. That is the point to add a classifier or judge, and, for actions that must never run, a policy.

FAQ

Does Sentry support AI agents?

Yes. Sentry traces agent runs, including LLM calls, tool executions, MCP interactions and handoffs, with inputs, outputs, tokens, cost and timing. This is useful for operational errors and performance. Behavioral failures still need evaluations or failure analysis that judges the agent's outcome.

Is Raindrop still "Sentry for AI"?

That was Raindrop's original YC launch title. Raindrop now monitors agents through tracing, plain-language Signals, ranked Issues, a Triage Agent and Experiments. It detects and investigates production behavior and sends alerts to Slack.

Why can an error tracker not catch agent failures on its own?

Because most agent failures raise nothing. An agent that gives a wrong answer, skips a step or declares a task done early completes every call successfully. An error tracker records it as a healthy run unless something - a code check, a classifier or an LLM judge - defines that outcome as a failure and reports it.

Get Started

Failproof AI is free to start. It finds recurring failure modes across agent sessions using code-based and LLM-based evaluations, groups the evidence into findings, and recommends fixes. Bring the eval suite you already have, alert the right owner when behavior drifts, and turn a tested fix into a policy that prevents the failure from recurring. See pricing for the tiers.

start free get a demo →

Sources

Checked against each vendor's own site and docs on 2026-09-14. Products change; if a detail here is out of date, tell us at support@befailproof.ai.