
AgentMarshal is a governance layer for autonomous AI agent fleets, built on Veea's Lobster Trap as a constitutive dependency. THE PROBLEM Businesses across industries are deploying autonomous AI agents with credentials, inboxes, and corporate cards. The capability landed before the governance layer did. One manipulated prompt can authorize a wire transfer, leak customer data, or commit a business to a contract the owner never approved. Real governance is four-dimensional: intent, vendor, category, cumulative spend. THE ARCHITECTURE Lobster Trap inspects every prompt with DPI β flagging injection patterns, obfuscation, computing a risk score. AgentMarshal consumes those signals and layers policy primitives on top: declared scope vs. detected intent, per-agent budgets, vendor allowlists, margin floors, approval thresholds. Every decision writes a full audit row. THE DEMO Cortez Roofing β 5-agent fleet in Phoenix. π’ GREEN β routine invoice approved. π‘ YELLOW β $14,800 quote at 28% margin against 35% floor. Escalates to Mike. One-click approval. π΄ RED β Comms Agent receives spoofed invoice from [email protected] with embedded prompt injection demanding new ACH routing. Lobster Trap fires risk_score 0.83 + injection + obfuscation. AgentMarshal blocks via block_prompt_injection. $12,000 attack blocked. Roofing is the example. The product is horizontal. DEFENSE-IN-DEPTH Lobster Trap is the inspection floor. AgentMarshal is the policy ceiling. Two layers, two jobs. One catches the conversation. The other catches the consequence. Tech: Next.js 14, TypeScript, YAML-driven policy engine (38/38 tests), Veea Lobster Trap (Go sidecar, MIT, unmodified), Ollama (local) / Groq llama-3.1-8b-instant (prod), SQLite, Fly.io. MIT licensed.
19 May 2026