diff --git a/AGENTS.md b/AGENTS.md
index 36bb5bf7..2fb00bb5 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -113,6 +113,16 @@ introducing a config-driven module — update `pkg/fx/app/full_test.go` so the
 new wiring is exercised. Add or adjust the storage/config variants in the test
 to cover the new graph shape.
 
+## Observability
+
+When a change adds or alters behavior an operator depends on, decide how they
+would see it working and see it fail, and add what is useful in the same PR:
+OpenTelemetry metrics or traces, entries in
+`docs/content/concepts/telemetry.md` and
+`docs/content/operator-guide/monitoring.md`, and an alert or dashboard panel in
+fil-forge/infra-central. Skip it for changes nothing operational depends on.
+"Adding a signal" in `docs/observability.md` says how to do each.
+
 ## Conventions
 
 - Go files: snake_case. Tests: testify (`assert`/`require`).
diff --git a/docs/observability.md b/docs/observability.md
new file mode 100644
index 00000000..689caf90
--- /dev/null
+++ b/docs/observability.md
@@ -0,0 +1,24 @@
+# Observability
+
+Contributor guidance for making a change observable. This file sits outside `docs/content/`, so it
+is not part of the published operator site.
+
+## Adding a signal
+
+When a change adds or alters behavior an operator depends on, decide how they would see it working
+and see it fail, and add that in the same PR:
+
+- **Metrics and traces.** Instrument with OpenTelemetry in the package's `telemetry.go` or
+  `*_metrics.go` (for example `pkg/curiopdp/metrics.go`, `lib/jobqueue/worker/telemetry.go`).
+  Prefer a signal that answers "is this working?", such as a failure count or the time of the last
+  success, over one that only measures volume. Keep attributes low-cardinality: no CIDs, DIDs or
+  other per-request values.
+- **Docs.** List each new metric in `docs/content/concepts/telemetry.md`, and in the "Key Metrics"
+  table of `docs/content/operator-guide/monitoring.md` if operators should watch it. If it shows a
+  failure an operator should act on, add a recommended alert with its severity to that page's
+  "Alerts to Configure" table, and its query to the PromQL block with the others.
+- **Forge's alerts and dashboards** live in fil-forge/infra-central, under
+  `terraform/envs/grafana/`. A new metric does nothing for Forge's on-call until a rule or panel
+  there reads it, so open a companion PR there, or say in this PR that one is needed.
+
+Skip this for changes nothing operational depends on.