diff --git a/.github/workflows/copilot-agent-analysis.md b/.github/workflows/copilot-agent-analysis.md
index 6e55e8d6037..ef8c734f553 100644
--- a/.github/workflows/copilot-agent-analysis.md
+++ b/.github/workflows/copilot-agent-analysis.md
@@ -18,6 +18,26 @@ permissions:
 engine: claude
 strict: true
 
+experiments:
+  output_format:
+    variants: [structured, prose]
+    description: "Test whether a prose-style discussion summary reduces AI credit consumption vs. the current table-centric structured format without sacrificing completeness."
+    hypothesis: "H0: no change in ai_credits_used. H1: prose format reduces ai_credits_used by >=15% while keeping empty_discussion_rate <=5%"
+    metric: ai_credits_used
+    secondary_metrics: [run_duration_seconds, output_length_chars]
+    guardrail_metrics:
+      - name: empty_discussion_rate
+        direction: min
+        threshold: 0.05
+    min_samples: 30
+    weight: [50, 50]
+    start_date: "2026-06-08"
+    issue: #aw_cop_fmt
+    analysis_type: t_test
+    tags: [cost-efficiency, output-quality, daily-report]
+    notify:
+      discussion: true
+
 network:
   allowed:
     - defaults
@@ -316,6 +336,7 @@ Create a **concise** discussion with your findings using the safe-outputs create
 
 **Discussion Title**: `Daily Copilot Agent Analysis - [DATE]`
 
+{{#if experiments.output_format == 'structured' }}
 **Concise Discussion Template**:
 ```markdown
 ### 🤖 Copilot Agent PR Analysis - [DATE]
@@ -401,6 +422,24 @@ The "Agent Task Texts" section should include a table showing all PRs created in
 - **Remove "Instruction File Changes"** section entirely
 - **Eliminate "Recommendations"** section - fold into "Key Insights" (1-2 bullets max)
 - **Remove verbose methodology** and historical context sections
+{{/if}}
+{{#if experiments.output_format == 'prose' }}
+**Prose Discussion Template**:
+```markdown
+### 🤖 Copilot Agent PR Analysis - [DATE]
+
+**Analysis Period**: Last 24 hours
+
+In the last 24 hours, Copilot agent created [count] PRs (`agent_prs_total`), of which [count] were merged ([percentage]% success rate, `agent_prs_merged`) with an average duration of [time] and [count] human comments per PR. [One sentence on 3-day trend only if success rate changed >10%, e.g. "Success rate improved from X% to Y% over the last 3 days." — otherwise omit.] [One sentence listing any notable PRs by number only if failures, closures, or PRs open >24h exist — otherwise omit.]
+
+- [Key insight 1: single most actionable observation — omit bullet entirely if nothing notable]
+- [Key insight 2: secondary pattern or trend worth flagging — omit bullet entirely if nothing notable]
+
+---
+
+_Generated by Copilot Agent Analysis (Run: [run_id])_
+```
+{{/if}}
 
 ## Important Guidelines