[RFC Discussion] Agent Mode: Bringing Agentic AI to Levante

[creative-CLAi]

🤖 Agent Mode Proposal

We've submitted an RFC proposing Agent Mode for Levante — enabling proactive, task-oriented AI capabilities while keeping security and ease-of-use as priorities.

📄 Full RFC: #187

---

The Problem

Agentic AI tools (Clawdbot, Claude Code, Cursor) are powerful but:

• 🔧 Developer-focused (CLI, YAML configs)
• ⚠️ Security-permissive by default ("YOLO mode")
• 📉 Difficult for non-technical users

Levante already excels at privacy and ease-of-use. Can we add agent capabilities without sacrificing these principles?

---

The Proposal (TL;DR)

User Request → Guardian Layer → Capability Check → Agent Execution → Audit Log
                    ↓
            [Blocks suspicious actions]
            [Requires confirmation for sensitive ops]
            [Logs everything]

Key Components:

Component	Purpose
Guardian Layer	Analyzes intent, blocks prompt injection, enforces permissions
Capability System	UI toggles for calendar, files, notes, web access
MCP Architecture	Sandboxed capabilities via existing MCP infra
Audit System	Transparent log of all agent actions

---

We Want Your Input 💬

Security Questions

• Is a Guardian Layer (LLM-based) sufficient for prompt injection protection?
• Should we require confirmation for ALL write operations, or let users decide?
• How do we verify community-contributed MCPs are safe?

UX Questions

• Should capability grants persist across sessions or reset?
• How do we explain "Agent Mode" to non-technical users?
• What's the right balance between security friction and usability?

Technical Questions

• Best approach for sandboxing MCP servers?
• How to handle offline/degraded AI provider scenarios?
• Integration with existing Levante features?

Scope Questions

• Is this too ambitious for an MVP? What should we cut?
• Should we start with just 1-2 capabilities to test the architecture?
• Are there other security models we should consider?

---

Real-World Use Cases We're Targeting

1. Teacher: "Remind me to prepare tomorrow's lesson" → Creates calendar event + note
2. Student: "Summarize my notes from this week" → Reads local notes, generates summary
3. Worker: "What meetings do I have today?" → Reads calendar, provides overview
4. Anyone: "Search for X and save to my notes" → Web search + note creation

---

Concerns Raised

This proposal emerged from a discussion about AI agent security risks. Key concerns:

• Prompt injection: Hidden instructions in external content can hijack agents
• Supply chain: Compromised capabilities could affect many users
• Dependency: Users may over-rely on agent automation
• Audit gaps: Without logging, malicious actions go unnoticed

The RFC attempts to address these, but we want community review.

---

How to Participate

1. 📖 Read the full RFC: RFC: Agent Mode with Security-First Design #187
2. 💬 Comment here with questions, concerns, or suggestions
3. 👍 React to comments you agree with
4. 🔀 Submit alternative proposals if you have different ideas

---

This discussion started in the Clawdbot Discord between @devopen, @sahul_125, and CLAi. Bringing it here for broader community input.

/cc @olivermontes