diff --git a/graph/decisions/agent-delegation-requires-careful-review-of-edge-cases-a3f466.md b/graph/decisions/agent-delegation-requires-careful-review-of-edge-cases-a3f466.md
new file mode 100644
index 00000000000..6e62ef0fc68
--- /dev/null
+++ b/graph/decisions/agent-delegation-requires-careful-review-of-edge-cases-a3f466.md
@@ -0,0 +1,20 @@
+---
+id: a3f46683-ef66-4fa9-8b44-eddca4d055a7
+type: decision
+title: "Agent delegation requires careful review of edge cases"
+tags: [cognitive-memory, agent-delegation, code-review, pattern]
+importance: 0.8
+confidence: 0.8
+created: "2026-02-13T17:39:40.881783+00:00"
+updated: "2026-02-13T17:39:40.881783+00:00"
+---
+
+When delegating implementation tasks to engineer agents, they reliably handle the core feature but miss edge cases in adjacent code paths. Two specific bugs found during Phase 1:
+
+1. Content preview staleness in update() - agent added content_preview to store/reindex but missed that update() also modifies content and needs to regenerate the preview.
+
+2. Content preview staleness in merge() - agent added merge command but called _update_index_entry() without content_preview after combining bodies.
+
+Pattern: When a feature touches a data field (like content_preview), every code path that modifies the underlying data must also update the derived field. Agents tend to handle the primary path but miss secondary mutation paths.
+
+Best practice: After agent delivers code, grep for all call sites of the modified methods and verify the new parameter/field is handled everywhere.