raindrop vs latitude

The Same Loop, Two Different Products

Raindrop and Latitude chase the same loop: trace the agent in production, find the failures, and hand the fix to a coding agent. The products behind that loop are built by different kinds of company, and it shows.

Raindrop, a YC W24 company in San Francisco, is a closed, hosted product with an emphasis on finish. A single failure in one run is a Stumble; the same stumble across many runs and users becomes an Issue, ranked by severity. Signals let you describe a behavior in plain language and get a classifier. A Triage Agent investigates root causes, and Experiments compare cohorts of already-logged production events. Its customer logos include Speak, Vercel, Clay and Framer.

Latitude, from Barcelona, is an open platform with an emphasis on the repair step. Its V2, released in June 2026 after several earlier product shapes, is agent observability only: OpenTelemetry tracing of sessions, tools, memory and cost; evaluations by rule, LLM judge or JavaScript; human annotation; monitors; and Agent Dispatch, which hands a failure to a coding agent that opens a pull request. The whole platform has been MIT-licensed since May 2026.

Side by Side

Two rows decide most choices here: where alerts go, and how the fix gets written. Everything else, including the price gap, follows from the kind of team you are.

One Broken Tool on a Friday Night

A backend engineer adds a required region parameter to the search_orders tool and ships it at 6pm on a Friday. The support agent keeps calling the tool the old way. Every call comes back with an error payload, the agent apologises to the customer and moves on, and nothing throws in your application. Here is how each fix loop handles the weekend.

1. Raindrop

   Each failed run is recorded as a Stumble; as it repeats across runs and users it becomes an Issue, ranked by severity. If you have a signal for "agent apologises instead of completing the task", a signal alert fires in Slack when its rate crosses the threshold, and the daily digest and any scheduled Agent Brief pick it up too. On Monday the Triage Agent looks for the root cause, and because anything you can do in Raindrop's UI a coding agent can do over MCP, you point Claude Code at the Issue and let it write the fix.

2. Latitude

   A rule-condition evaluation on tool results, which Latitude's docs call "deterministic checks, free and instant", flags each completed session with the error, subject to the evaluation's sampling. A monitor fires in-app, by email and in Slack. If Agent Dispatch is set to trigger on monitor incidents, it hands the failure to Claude Code or Cursor, which investigates through Latitude's MCP and opens a PR adding the parameter. On Monday there is a pull request waiting for review.

Raindrop gives you a better-organised picture of the weekend: which issue, how many users, how severe, in the channel your team already lives in. Latitude gives you a draft of the fix. Whether you would rather wake up to a ranked issue or a PR depends on how much you trust an unattended coding agent with your tool code, and Latitude lets you set per-project guardrails, mute signals you do not want acted on, and review every PR before it merges.

Writing the Checks

Raindrop's checks start from a sentence. You describe a behavior on the Signals page, Raindrop's agent reads your traces, writes the rules and builds a classifier, and you mark a sample of matches as Match or Not Match until it is right. You never label every event. The result is a yes/no signal with a rate you can chart, alert on and use in Experiments. For a team with no evaluation experience, this is the fastest way in either product to go from "users are annoyed about something" to a number.

Latitude's checks start from a choice of three detectors, in its docs' words: "Set of conditions: deterministic checks, free and instant", "LLM as judge: describe the behavior and let an LLM decide", and "Custom script: JavaScript for anything the other two can't express". Evaluations can be generated from signals, and Latitude keeps generated ones aligned to human annotations over time; one you write yourself "runs exactly as written". The Friday failure above is a job for the first kind: a rule on tool results that costs nothing to run on every sampled session.

Where Each One Is Stronger

Raindrop

• Triage polish. Stumbles rolling up into severity-ranked Issues turn a pile of failed runs into a short list, and the Triage Agent does the first pass of root-cause work. It is the most finished part of either product.
• Checks from plain language. Signals with a labelling loop get a non-specialist to a working classifier quickly.
• Evidence from production. Experiments compare logged cohorts before and after a change, with confidence intervals; Latitude does not publish an equivalent.
• Customer proof. Well-known AI-native customers such as Speak, Vercel and Clay, and a public case study with a number in it.

Latitude

• Ownership. The whole platform is MIT, and self-hosting it is free.
• Price of entry. $99 a month for Pro, and unlimited seats from the free tier up.
• Alerts beyond Slack. Monitors reach email and the app as well as Slack.
• Hands-off repair. Agent Dispatch can trigger on its own and deliver a pull request, not just a ticket.
• Data location. A cloud in AWS eu-central-1 suits EU teams without a self-host.

Alerts and On-Call

Raindrop's alerting docs are clear: signal alerts, the daily digest and Agent Briefs "are delivered to Slack". No email, webhook or paging integration is documented. For a team whose entire operation runs in Slack, that is the right default and one less integration to maintain. For a team whose on-call rota runs through a paging tool, it means someone has to bridge Slack to it.

Latitude's monitors notify in-app, by email and in Slack, and Agent Dispatch can send to Linear or any webhook, which is a workable route into other systems. Neither documents a native PagerDuty integration.

Proving the Fix Worked

This is where Raindrop is clearly stronger. Experiments compare a baseline and an experiment cohort of already-logged production events, defined by model, feature flag, property, tool or date range, and report each signal's change with a 95% confidence interval and a p-value, labelled likely better, likely worse or mixed. Experiments "don't run or replay your agent"; they read what already happened. With enough users, that is the most honest evidence a fix can get. With too few, two small cohorts make a noisy comparison, and the docs state no minimum to plan around.

Latitude builds datasets from real traces for replays and regression tests, so you can re-run the Friday failures against the patched agent before merging the PR. That works at any volume and catches the regression before a customer sees it, but it tells you the fix works on the cases you kept, not how it changed behavior across everyone.

Price, Licence and Where It Runs

Raindrop's bill is easy to model: at 100,000 events a month, Pro is $299 + $300 = $599. Latitude's is not, because its pricing page asks "What is a credit and how are they calculated?" without answering it there. The headline gap, $299 against $99, is real at the bottom of the range; what happens at your volume depends on a unit you have to ask about.

The licence gap is not a matter of degree. Raindrop open-sources Workshop, an MIT local debugger and eval runner, and keeps the platform closed. Latitude open-sources the platform itself, and its self-host docs say every option "deploys the same images" as the cloud. If owning the system matters more than polish, that settles it.