diff --git a/AGENT_LOOP_IMPROVEMENTS.md b/AGENT_LOOP_IMPROVEMENTS.md
index 3ca571a7d..5eb70a4e6 100644
--- a/AGENT_LOOP_IMPROVEMENTS.md
+++ b/AGENT_LOOP_IMPROVEMENTS.md
@@ -1,23 +1,51 @@
+Here is the complete, updated **Agent Loop Improvements Plan** markdown file, incorporating the refined architecture for Phase 3.
+
+---
+
 # Agent Loop Improvements Plan
 
 This document outlines a series of tasks to improve the main loop of the agentic stack (`AgentLoop` in `pkg/agent/loop.go`), focusing on Performance, Reliability, Architecture, and Features.
 
 ## Phase 1: Architecture & Maintainability (Refactoring)
-- [x] **Extract LLM Call & Retry Logic:** Refactor `runLLMIteration` by moving the LLM calling, fallback chain, and context window retry logic into a dedicated method (e.g., `executeLLMWithRetry`).
-- [x] **Extract Tool Execution Logic:** Refactor `runLLMIteration` by moving the parallel tool execution (`sync.WaitGroup`), channel routing, and async callbacks into a dedicated method (e.g., `executeToolBatch`).
-- [x] **State Machine / Explicit Flow:** Clean up the main loop logic to reduce nested `if/for` blocks and make the transition between generating, executing tools, and compressing context more explicit.
+
+* [x] **Extract LLM Call & Retry Logic:** Refactor `runLLMIteration` by moving the LLM calling, fallback chain, and context window retry logic into a dedicated method (e.g., `executeLLMWithRetry`).
+* [x] **Extract Tool Execution Logic:** Refactor `runLLMIteration` by moving the parallel tool execution (`sync.WaitGroup`), channel routing, and async callbacks into a dedicated method (e.g., `executeToolBatch`).
+* [x] **State Machine / Explicit Flow:** Clean up the main loop logic to reduce nested `if/for` blocks and make the transition between generating, executing tools, and compressing context more explicit.
 
 ## Phase 2: Reliability & Error Handling
-- [x] **Graceful Recovery on Tool Panic:** Add `defer recover()` inside the parallel tool execution goroutines to prevent a panicked tool from crashing the entire `AgentLoop`. Return the panic as an error string to the LLM.
-- [x] **Exponential Backoff:** Replace the linear backoff in LLM retries (`time.Duration(retry+1) * 5 * time.Second`) with exponential backoff and jitter to better handle rate limits.
-- [ ] **Granular Error Classification:** Update `LLMProvider` interfaces to return structured, typed errors (e.g., `providers.ErrContextLengthExceeded`) instead of relying on fragile string matching.
+
+* [x] **Graceful Recovery on Tool Panic:** Add `defer recover()` inside the parallel tool execution goroutines to prevent a panicked tool from crashing the entire `AgentLoop`. Return the panic as an error string to the LLM.
+* [x] **Exponential Backoff:** Replace the linear backoff in LLM retries (`time.Duration(retry+1) * 5 * time.Second`) with exponential backoff and jitter to better handle rate limits.
+* [ ] **Granular Error Classification:** Update `LLMProvider` interfaces to return structured, typed errors (e.g., `providers.ErrContextLengthExceeded`) instead of relying on fragile string matching.
 
 ## Phase 3: Performance & Latency
-- [x] **Concurrent Message Processing:** Evaluate introducing a worker pool or goroutines to process independent user requests concurrently without blocking the entire agent instance.
-- [ ] **Background Summarization:** Offload `maybeSummarize` and context compression to a background worker to unblock the main thread and respond to the user faster.
-- [ ] **Streaming Responses:** Implement streaming LLM token generation directly to the `bus.PublishOutbound` instead of waiting for full generation.
+
+* [x] **Concurrent Message Processing:** Evaluate introducing a worker pool or goroutines to process independent user requests concurrently without blocking the entire agent instance.
+* [ ] **Background Summarization:** Offload `maybeSummarize` and context compression to an asynchronous worker. Instead of blocking the main thread, the worker outputs its execution trace and compressed context to a structured `/logs/{session}/{subagent}/` directory to provide a persistent audit trail while keeping the loop responsive.
+* [ ] **Streaming Responses:** Implement streaming LLM token generation directly to the `bus.PublishOutbound` instead of waiting for full generation.
 
 ## Phase 4: Features
-- [ ] **Human-in-the-Loop:** Introduce an approval prompt state for high-risk tools (e.g., SQL execution) that pauses the loop until a user explicitly replies "Yes/No".
-- [ ] **Background / Long-Running Tasks:** Implement tools that allow the LLM to run slow operations in the background, releasing the main loop and notifying the user asynchronously upon completion.
-- [ ] **Multi-Agent Orchestration:** Create a Supervisor Loop where an agent can delegate tasks to other `AgentInstance`s and synthesize their results.
+
+* [ ] **Human-in-the-Loop:** Introduce an approval prompt state for high-risk tools (e.g., SQL execution) that pauses the loop until a user explicitly replies "Yes/No".
+* [ ] **Background / Long-Running Tasks:** Implement tools that allow the LLM to run slow operations in the background, releasing the main loop and notifying the user asynchronously upon completion.
+* [ ] **Multi-Agent Orchestration:** Create a Supervisor Loop where an agent can delegate tasks to other `AgentInstance`s and synthesize their results.
+
+---
+
+### Implementation Note: Subagent Logging Structure
+
+To ensure observability for Phase 3's background tasks, the following directory convention is recommended:
+
+```text
+/logs
+└── {session_id}/
+    └── summarizer/
+        ├── trace.log          # Real-time workflow steps
+        ├── input_context.json  # Snapshot of data before compression
+        └── result_summary.md  # Final output to be injected into main state
+
+```
+
+---
+
+**Would you like me to focus on the Go implementation for the `BackgroundSummarizer` or the `Granular Error Classification` next?**