diff --git a/docs/monitoring/ADAPTIVE_POLLING.md b/docs/monitoring/ADAPTIVE_POLLING.md
new file mode 100644
index 0000000..f2774db
--- /dev/null
+++ b/docs/monitoring/ADAPTIVE_POLLING.md
@@ -0,0 +1,91 @@
+# Adaptive Polling Architecture
+
+## Overview
+Phase 2 introduces a scheduler that adapts poll cadence based on freshness, errors, and workload. The goal is to prioritize stale or changing instances while backing off on healthy, idle targets.
+
+```
+┌──────────┐      ┌──────────────┐      ┌──────────────┐      ┌─────────────┐
+│ PollLoop │─────▶│  Scheduler   │─────▶│ Priority Q   │─────▶│ TaskWorkers │
+└──────────┘      └──────────────┘      └──────────────┘      └─────────────┘
+        ▲                   │                    │                    │
+        │                   └─────► Staleness, metrics, circuit breaker feedback ────┘
+```
+
+- **Scheduler** computes `ScheduledTask` entries using adaptive intervals.
+- **Task queue** is a min-heap keyed by `NextRun`; only due tasks execute.
+- **Workers** execute tasks, capture outcomes, reschedule via scheduler or backoff logic.
+
+## Key Components
+
+| Component             | File                                      | Responsibility                                               |
+|-----------------------|-------------------------------------------|--------------------------------------------------------------|
+| Scheduler             | `internal/monitoring/scheduler.go`        | Calculates adaptive intervals per instance.                 |
+| Staleness tracker     | `internal/monitoring/staleness_tracker.go`| Maintains freshness metadata and scores.                    |
+| Priority queue        | `internal/monitoring/task_queue.go`       | Orders `ScheduledTask` items by due time + priority.        |
+| Circuit breaker       | `internal/monitoring/circuit_breaker.go`  | Trips on repeated failures, preventing hot loops.           |
+| Backoff               | `internal/monitoring/backoff.go`          | Exponential retry delays with jitter.                       |
+| Workers               | `internal/monitoring/monitor.go`          | Pop tasks, execute pollers, reschedule or dead-letter.      |
+
+## Configuration
+
+Environment variables (default in `internal/config/config.go`):
+
+| Variable                            | Default | Description                                      |
+|-------------------------------------|---------|--------------------------------------------------|
+| `ADAPTIVE_POLLING_ENABLED`          | false   | Feature flag for adaptive scheduler.             |
+| `ADAPTIVE_POLLING_BASE_INTERVAL`    | 10s     | Target cadence when system is healthy.           |
+| `ADAPTIVE_POLLING_MIN_INTERVAL`     | 5s      | Lower bound (active instances).                  |
+| `ADAPTIVE_POLLING_MAX_INTERVAL`     | 5m      | Upper bound (idle instances).                    |
+
+All settings persist in `system.json` and respond to environment overrides.
+
+## Metrics
+
+Exposed via Prometheus (`:9091/metrics`):
+
+| Metric                                   | Type      | Labels                                | Description                                |
+|------------------------------------------|-----------|---------------------------------------|--------------------------------------------|
+| `pulse_monitor_poll_total`               | counter   | `instance_type`, `instance`, `result` | Overall poll attempts (success/error).     |
+| `pulse_monitor_poll_duration_seconds`    | histogram | `instance_type`, `instance`           | Poll latency per instance.                 |
+| `pulse_monitor_poll_staleness_seconds`   | gauge     | `instance_type`, `instance`           | Age since last success (0 on success).     |
+| `pulse_monitor_poll_queue_depth`         | gauge     | —                                     | Size of priority queue.                    |
+| `pulse_monitor_poll_inflight`            | gauge     | `instance_type`                       | Concurrent tasks per type.                 |
+
+## Circuit Breaker & Backoff
+
+| State       | Trigger                                     | Recovery                                   |
+|-------------|---------------------------------------------|--------------------------------------------|
+| **Closed**  | Default. Failures counted.                  | —                                          |
+| **Open**    | ≥3 consecutive failures. Poll suppressed.   | Exponential delay (max 5 min).             |
+| **Half-open**| Retry window elapsed. Limited re-attempt. | Success ⇒ closed. Failure ⇒ open.         |
+
+Backoff configuration:
+
+- Initial delay: 5 s
+- Multiplier: x2 per failure
+- Jitter: ±20 %
+- Max delay: 5 minutes
+- After 5 transient failures or any permanent failure, task moves to dead-letter queue for operator action.
+
+## Dead-Letter Queue
+
+Dead-letter entries are kept in memory (same `TaskQueue` structure) with a 30 min recheck interval. Operators should inspect logs for `Routing task to dead-letter queue` messages. Future work (Task 8) will add API surfaces for inspection.
+
+## Operational Guidance
+
+1. **Enable adaptive polling**: set `ADAPTIVE_POLLING_ENABLED=true` via UI or environment overrides, then restart hot-dev (`scripts/hot-dev.sh`).
+2. **Monitor metrics** to ensure queue depth and staleness remain within SLA. Configure alerting on `poll_staleness_seconds` and `poll_queue_depth`.
+3. **Review dead-letter logs** for persistent failures; resolve underlying connectivity or auth issues before re-enabling.
+
+## Rollout Plan
+
+1. **Dev/QA**: Run hot-dev with feature flag enabled; observe metrics and logs for several cycles.
+2. **Staged deploy**: Enable flag on a subset of clusters; monitor queue depth (<50) and staleness (<45 s).
+3. **Full rollout**: Toggle flag globally once metrics are stable; document any overrides in release notes.
+4. **Post-launch**: Add Grafana panels for queue depth & staleness; alert on circuit breaker trips (future API work).
+
+## Known Follow-ups
+
+- Task 8: expose scheduler health & dead-letter statistics via API and UI panels.
+- Task 9: add dedicated unit/integration harness for the scheduler & workers.
+