A production agent platform with 97 authorized business capabilities and a sandbox where models can analyze live data without directly mutating it.
Giving an LLM real tools against a real ERP is where most agent projects quietly break. Ninety-odd tools means ninety-odd places to forget a permission check. And the moment you let the model run code so it can read across several endpoints at once, you have handed it the ability to write too.
I unified in-app and external-agent tools behind one capability registry and authorization boundary. For multi-step analysis, LLM-authored TypeScript runs in a QuickJS/WASM isolate with per-user read-only bindings, budgets, and deadlines. Writes are structurally unavailable inside the sandbox; proposed actions return to the user for confirmation.
Live in production. Replacing the failing async runtime made multi-read code execution reliable while preserving a portable, pure-WASM deployment.
My Role: AI Engineer & Architect