Use case
Give agent runs a structured completion inbox
Long-running agents often finish after you have moved to another task. A small HTTP report gives each run a durable outcome, exact counters and a safe path back to its artifacts.
Updated 6 min read
Send the outcome, not the whole transcript
- Use a clear title for completed, failed, skipped or degraded outcomes.
- Put bounded counters, run ID and modelled quality signals in metadata.
- Link to the controlled artifact or run view instead of copying sensitive output.
Use stable IDs across retries and related events
The producer's run ID belongs in metadata and can also anchor the deduplication key. Add an outcome suffix if the same run legitimately reports more than one state transition.
Choose a source per agent workflow
This keeps permissions revocable and filters predictable without pretending every run is a different integration.
Use tags for durable categories
Classify research, publication or maintenance runs independently from their changing result values.
Frequently asked questions
- Does the producer need a special integration?
- No. The examples use an authenticated HTTPS request, so any script, service or visual workflow that can send HTTP can use the same contract.
- Should credentials be included in notification data?
- No. Keep credentials in the producer's secret store and send only the context needed to understand or locate the event.
Keep exploring
Put the pattern to work
Existing accounts can open the inbox and send their next event.