diff --git a/agent.md b/agent.md
new file mode 100644
index 0000000..1a5209b
--- /dev/null
+++ b/agent.md
@@ -0,0 +1,267 @@
+# agent.md
+
+**Project Codename:** IntentOS
+**Category:** Intent-Native Productivity System
+**Primary Stack:** Next.js (App Router), Tambo SDK, Gemini 3
+**Execution Mode:** Solo / Autonomous Agent Compatible
+
+## 1. Mission Statement
+
+You are building **IntentOS**, a next-generation productivity system where:
+
+> Users express goals in natural language, and the system dynamically renders interactive workflows instead of static screens.
+
+This is **not** a chat app.
+Chat is the control surface.
+UI is a side-effect of intent reasoning.
+
+Your objective is to ship a **fully working, judge-ready demo** that demonstrates this paradigm clearly within hackathon constraints.
+
+## 2. Core Design Principles (Non-Negotiable)
+
+1. **Intent > Screens**
+
+   - No fixed dashboards
+   - No static workflows
+   - UI components appear only when required by intent
+
+2. **Chat as an OS Bus**
+
+   - All interaction flows through Tambo’s message system
+   - Threads represent *one intent lifecycle*
+
+3. **Minimal Surface, Maximum Depth**
+
+   - One killer end-to-end demo flow
+   - Avoid feature sprawl
+
+4. **Leverage Existing Tambo Primitives**
+
+   - Do not rebuild chat, editors, or orchestration logic
+   - Compose, don’t replace
+
+## 3. High-Level Architecture
+
+```
+User Input (Text / Voice / Files)
+        ↓
+Tambo Message Pipeline
+        ↓
+Gemini 3 (Intent Classification + Planning)
+        ↓
+Structured Intent JSON
+        ↓
+Elicitation (if needed)
+        ↓
+Dynamic Component Rendering
+        ↓
+Interactive Workflow Execution
+```
+
+## 4. Canonical Intent Contract (CRITICAL)
+
+All intents must resolve into a structured object.
+
+### Intent JSON Schema (Baseline)
+
+```json
+{
+  "intent_id": "uuid",
+  "intent_type": "goal_planning | analysis | tracking | decision_support",
+  "summary": "string",
+  "confidence": 0.0,
+  "required_components": [
+    "elicitation",
+    "timeline",
+    "task_list",
+    "chart"
+  ],
+  "elicitation_schema": {
+    "fields": [
+      {
+        "key": "time_horizon",
+        "type": "string",
+        "required": true
+      }
+    ]
+  },
+  "workflow_state": "draft | active | completed"
+}
+```
+
+This schema is the **bridge between Gemini and Tambo UI rendering**.
+
+## 5. Agent Responsibilities (Task Breakdown)
+
+### A. Intent Understanding Layer (Gemini)
+
+**Objective:** Convert free-form user input into structured intent.
+
+Tasks:
+
+- Write a **system prompt** for Gemini that:
+  - Classifies intent
+  - Determines missing information
+  - Outputs valid JSON only
+- Ensure deterministic structure
+- Reject vague or multi-intent inputs with clarification requests
+
+Deliverable:
+
+- A single prompt file or inline system prompt
+- Sample intent outputs stored for demo reliability
+
+### B. Elicitation Layer (Tambo Native)
+
+**Objective:** Collect missing parameters before execution.
+
+Use:
+
+- `ElicitationUI`
+
+Tasks:
+
+- Dynamically render questions from `elicitation_schema`
+- Support:
+  - Single value fields
+  - Multi-entry fields
+- Lock intent once elicitation completes
+
+Deliverable:
+
+- Clean elicitation → execution transition
+- No dead-ends or infinite loops
+
+### C. Dynamic Workflow Rendering
+
+**Objective:** Render UI components based on intent needs.
+
+Allowed Components:
+
+- `ElicitationUI`
+- `DataCard`
+- `Graph`
+- Markdown blocks with embedded components
+
+Rules:
+
+- No page navigation
+- No hardcoded workflows
+- Rendering must be driven by intent JSON
+
+Deliverable:
+
+- One flagship workflow fully implemented
+
+### D. Productivity Primitives
+
+Implement **only what supports the demo intent**.
+
+Examples:
+
+- Task checklist with completion state
+- Timeline visualization
+- Progress or status indicator
+
+State Handling:
+
+- State may be local or ephemeral
+- Persistence is optional, clarity is not
+
+### E. Thread & Memory Semantics
+
+Use:
+
+- `ThreadHistory`
+
+Rules:
+
+- One thread = one intent lifecycle
+- Thread title = intent summary
+- Thread history doubles as “intent archive”
+
+### F. MCP Integration (Signal, Not Depth)
+
+Use:
+
+- `McpPromptButton`
+- `McpResourceButton`
+- `McpConfigModal`
+
+Tasks:
+
+- Show MCP affordances
+- Demonstrate at least one injected prompt or resource
+- No real server required
+
+Purpose:
+
+- Signal extensibility
+- Impress judges with forward compatibility
+
+## 6. Non-Goals (Explicitly Out of Scope)
+
+Do NOT implement:
+
+- Authentication logic
+- User settings
+- Notifications
+- Collaboration
+- Theming systems
+- Multi-intent orchestration
+
+If it does not strengthen the **intent → workflow** story, it is a liability.
+
+## 7. Demo Flow (Must Be Perfect)
+
+### Required Demo Scenario
+
+> “Create a 30-day internship preparation plan”
+
+Steps:
+
+1. User enters intent via chat or voice
+2. Gemini returns structured intent
+3. Elicitation collects missing constraints
+4. System renders:
+   - Timeline
+   - Task checklist
+   - Progress visualization
+5. User interacts with tasks
+6. System updates state and suggests next steps
+
+This flow must be:
+
+- Fast
+- Visually clear
+- Impossible to confuse with a normal chatbot
+
+## 8. Judging Alignment Checklist
+
+- **Tambo Usage:** Central, not decorative
+- **Generative UI:** Clearly visible
+- **Innovation:** Intent-native paradigm obvious in first 30 seconds
+- **Technical Execution:** Clean, stable, explainable
+- **Narrative:** “We design intent, not screens”
+
+## 9. Success Criteria
+
+The project is successful if:
+
+- A judge can explain IntentOS after one demo
+- The UI adapts visibly to user intent
+- The system feels like infrastructure, not an app
+- The codebase looks intentional, not hacked together
+
+## 10. Final Instruction to Agent
+
+Optimize for:
+
+- Clarity over completeness
+- Depth over breadth
+- Systems thinking over UI polish
+
+This is not a feature contest.
+This is a **category definition exercise**.
+
+Build accordingly.