How a dispatch works
Latitude sends the production context directly to your coding agent. No one has to copy traces into a ticket.
Paste this into Cursor, Claude Code, Codex, or any coding agent in the repository that produced these traces.
- const result = await tool.call(input)+ const result = await withRetry(() => tool.call(input), { max: 2 })
Give your coding agent the full production context
A dispatch gives your coding agent the Signal, trend, sample traces, and MCP access to investigate the underlying spans. If it finds a concrete repo-level root cause, it implements the fix and opens a PR. If it doesn’t, it returns an investigation summary without changing your code.
What lands in the prompt
- signalname, source, priority, and the description of the failure mode
- trendoccurrences, time window, and baseline per hour
- tracesup to 5 failing trace IDs
- conversationsup to 2 full rendered excerpts
- tagsthe top tags the failing traces share
- feedbackthe latest human annotation or eval result
- linka deep link into the Signal in Latitude
- mcpthe agent queries Latitude directly: getSignal, listSignalTraces, getTrace, querySpans
What wakes the agent
Pick which of these dispatch. The rest stay quiet.
- New SignalA failure pattern is auto-discovered in production traffic
- Escalating SignalThe last 24h of occurrences beats the same hours the week before
- Regressed SignalA Signal you resolved starts recurring
- Monitor incidentA saved search or raw traffic metric crosses its threshold
Or send one yourself, any time, with the Send to button on a Signal.
Where it dispatches to
Connect a target once in project settings.
- Cursor
- Claude Code
- Linear
- Your own webhook
Built in the open
Agent Dispatch is part of Latitude’s open-source platform. Inspect the implementation, self-host it, or connect through standard OpenTelemetry.
Trusted by teams running agents in production
Questions engineers actually ask
Does Latitude run the coding agent?
No. Latitude is the trigger and the context provider. The agent runs on Cursor or Anthropic infrastructure — or your own, through the webhook target — in your environment, with your credentials, against your repository.
Can it change or merge code without me?
It opens a pull request. Nothing lands until you review and merge it, exactly like a PR from a teammate.
What happens when the failure is not a code bug?
The agent is explicitly instructed not to make speculative changes. If it cannot find a repo-level root cause it reports back with what it found and opens no PR.
Can I trigger a dispatch myself?
Yes. The Send to button on any Signal hands the same context to Cursor or Claude Code on demand, or copies the assembled prompt so you can paste it into Codex, OpenCode, or whatever you run locally.
What do I need before this works?
Traces flowing into Latitude — our SDK, or standard OTLP/HTTP straight to https://ingest.latitude.so/v1/traces — and one integration connected in project settings. Getting traces in takes about five minutes.
What does the agent get over MCP, exactly?
A remote, OAuth-authenticated MCP server at api.latitude.so/v1/mcp. You connect it once in your Cursor environment or Claude routine — Latitude never mints or forwards MCP credentials at dispatch time. The blast radius of a leaked dispatch credential is "can start agent runs on your platform account", not "can read Latitude data".
What happens when a dispatch fails?
Four attempts with exponential backoff from 30 seconds, but only for rate-limit and transport errors. Auth and config failures fail fast rather than retrying into a wall, and every attempt lands in the dispatch history with its error category.
Stop relaying context by hand
Get traces into Latitude in about five minutes, connect your coding agent, and let the next failure fix itself while you review the diff.
Free plan, no credit card required.