diff --git a/docs/operation-modes.md b/docs/operation-modes.md
new file mode 100644
index 000000000..78e988c67
--- /dev/null
+++ b/docs/operation-modes.md
@@ -0,0 +1,140 @@
+# Indexer Agent Operation Modes
+
+This document explains the internal workings of the indexer agent, focusing on allocation management, the reconciliation loop, and how indexing rules interact with allocation actions.
+
+## Allocation Management Modes
+
+The agent supports three allocation management modes, configured via `--allocation-management`:
+
+| Mode | Behavior |
+|------|----------|
+| **AUTO** (default) | Allocation decisions are automatically approved and executed |
+| **MANUAL** | Reconciliation is completely skipped; all actions must be manually created via CLI |
+| **OVERSIGHT** | Actions are queued but require manual approval before execution |
+
+## The Reconciliation Loop
+
+The reconciliation loop runs continuously with two polling intervals:
+
+- **Small interval**: `pollingInterval` - for frequently changing data
+- **Large interval**: `pollingInterval * 5` - for stable data
+
+### Data Streams
+
+The loop fetches and maintains:
+
+1. **Current epoch number** (large interval)
+2. **Max allocation duration** (large interval)
+3. **Indexing rules** (small interval)
+4. **Active deployments on Graph Node** (large interval, AUTO mode only)
+5. **Network deployments** (small interval)
+6. **Active allocations** (small interval)
+7. **Recently closed allocations** (small interval)
+
+### Two-Phase Process
+
+The reconciliation consists of two distinct phases:
+
+#### Phase 1: Deployment Evaluation
+
+**What it does**: Determines which deployments the indexer *should* be allocated to based on indexing rules.
+
+- Takes all network deployments from the network subgraph
+- Matches each deployment against indexing rules (deployment-specific or global)
+- Outputs allocation decisions with `toAllocate: true/false` for each deployment
+
+**When it runs**: Every `pollingInterval * 5`, but **only in AUTO/OVERSIGHT mode**. Skipped entirely in MANUAL mode.
+
+**Decision basis options**:
+- `always` - allocate
+- `never` - don't allocate
+- `offchain` - index but don't allocate
+- `rules` - evaluate against thresholds (`minStake`, `minSignal`, `minAverageQueryFees`)
+
+#### Phase 2: Allocation Reconciliation
+
+**What it does**: Compares desired state (from deployment evaluation) against actual on-chain state and queues actions to resolve differences.
+
+| Condition | Action Queued |
+|-----------|---------------|
+| `toAllocate=true` + no active allocation | ALLOCATE |
+| `toAllocate=false` + active allocation exists | UNALLOCATE |
+| `toAllocate=true` + allocation expiring | REALLOCATE |
+
+**Expiration check**: `currentEpoch >= createdAtEpoch + allocationLifetime`
+
+**When it runs**: After deployment evaluation, **only in AUTO/OVERSIGHT mode**. Skipped in MANUAL mode.
+
+### Visual Flow
+
+```
+AUTO/OVERSIGHT MODE:
+┌─────────────────────┐     ┌──────────────────────────┐     ┌─────────────┐
+│ Deployment          │ --> │ Allocation               │ --> │ Action      │
+│ Evaluation          │     │ Reconciliation           │     │ Executor    │
+│                     │     │                          │     │             │
+│ "What SHOULD we     │     │ "What do we need to DO   │     │ Execute     │
+│  allocate to?"      │     │  to match desired state?"│     │ on-chain    │
+└─────────────────────┘     └──────────────────────────┘     └─────────────┘
+
+MANUAL MODE:
+┌─────────────────────┐     ┌──────────────────────────┐     ┌─────────────┐
+│ Deployment          │     │ Allocation               │     │ Action      │
+│ Evaluation          │     │ Reconciliation           │     │ Executor    │
+│                     │     │                          │     │             │
+│ SKIPPED             │     │ SKIPPED                  │     │ Still runs! │
+└─────────────────────┘     └──────────────────────────┘     └─────────────┘
+                                                                   ^
+                                                                   |
+                                                          Manual CLI actions
+                                                          (graph indexer actions queue)
+```
+
+## Indexing Rules
+
+### Purpose
+
+Rules serve two purposes:
+
+1. **Allocation decisions** (AUTO/OVERSIGHT only): Determine whether to allocate via `decisionBasis` and threshold fields
+2. **Deployment management** (ALL modes): Control what to index on Graph Node, provide defaults for manual actions
+
+### Management
+
+Rules are primarily managed through:
+
+- **CLI**: `graph indexer rules set/delete/clear/get`
+- **GraphQL API**: Direct mutations to the indexer management server
+
+### Automatic Rule Modifications
+
+**Important**: Allocation actions automatically modify rules to maintain consistency:
+
+| Action | Rule Modification |
+|--------|------------------|
+| **ALLOCATE** | Creates rule with `decisionBasis: ALWAYS` if no matching rule exists |
+| **UNALLOCATE** | **Always** sets rule to `decisionBasis: NEVER` |
+| **REALLOCATE** | Creates rule with `decisionBasis: ALWAYS` if no matching rule exists |
+
+This means:
+- After unallocating, you must manually change the rule back to `ALWAYS` or `RULES` if you want to allocate again
+- The agent won't fight against manual allocation actions (it creates rules to match)
+
+## Safety Mechanisms
+
+The agent includes several safety features:
+
+1. **Health check**: Won't allocate to deployments with "failed" health status if rule has `safety=true`
+2. **Zero POI safety**: Won't reallocate if previous allocation closed with zero POI and safety is enabled
+3. **Approved actions check**: Skips reconciliation if there are already pending approved actions (prevents conflicts)
+4. **Network subgraph protection**: Never auto-allocated unless explicitly enabled via `--allocate-on-network-subgraph`
+
+## Action Execution
+
+Actions flow through the action queue regardless of mode:
+
+1. **AUTO mode**: Actions are queued with `APPROVED` status and executed automatically
+2. **OVERSIGHT mode**: Actions are queued with `QUEUED` status, require manual approval
+3. **MANUAL mode**: No automatic actions; user queues actions via CLI with desired status
+
+The action executor runs in all modes - in MANUAL mode it simply has no auto-generated actions to process.