Under the Hood — New AI Agent Engine

This is the kind of update you shouldn't notice — and that's the point. We've completely rebuilt the AI engine that powers every text conversation on Fedna AI. Same features, same interface, fundamentally better foundation.

What changed

The custom AI agent runner that powered WhatsApp, web, and Co-Pilot conversations has been replaced with a production-grade SDK-based engine. Every conversation — customer-facing and internal — now runs on the new stack.

What you get out of it

Structured tool calling

When the AI calls a skill, generates an image, searches knowledge, or takes any action — it now uses structured, validated tool calls. Less ambiguity, fewer misfires, more reliable skill execution.

Sub-agent orchestration

The AI can now delegate work to specialised sub-agents under the hood — one to handle a knowledge lookup, another to process a skill call, another to format the response. This means faster, more accurate answers on complex queries.

Self-hosted AI tracing

Every AI decision — which model was called, what tools fired, how long it took, what the reasoning chain looked like — is now traced and observable via our self-hosted tracing infrastructure. This powers the debugging, monitoring, and quality improvements we ship going forward.

Better model resolution

The engine validates every model choice against the AI gateway in real time. No more silent fallbacks to an unexpected model if something is misconfigured.

Zero-downtime migration

We rolled this out in phases over multiple deploys with feature flags — the old engine and new engine ran side by side until we validated every path. No conversations were dropped, no features were lost.

What it enables next

The new engine is the foundation for everything coming next — multi-step reasoning, richer Co-Pilot capabilities, smarter agent-to-agent handoffs, and tighter integration between voice and text agents. This was the prerequisite.