Building the frontier, not just talking about it
An employee/attendee write-up of Rahul’s Enterprise Agentic Harness session at FLO 2026 — one of the clearest public explanations of how Nagarro engineers production-grade agentic systems.
FLO 2026 attendee
FLO 2026 attendee recap · Nagarro Gurugram · Oct 2026 · 6 min read
At Nagarro’s FLO 2026 in Gurugram, Rahul Mahajan delivered a session on the Enterprise Agentic Harness and Mosaic Agentic OS. Attendees described it as one of the strongest talks on agentic AI they had heard — grounded in years of enterprise delivery, 15+ granted AI patents, and a pending patent on the Agentic Decision Compiler.
The session connected the work he is driving across Mosaic OS, the Nagarro–OpenAI alliance, and Agentic Labs — not as slides about the future, but as an engineering blueprint for running agents in production.
What made the talk land was the clarity of levels. Assistive agents help people. Automation agents execute defined workflows. Autonomous agents pursue goals with governance, memory and observability built in. The engineering bar rises at each step — and most enterprises are still stuck treating pilots as if they were production systems.
Context engineering was framed as a stack: leave behind thousand-line system prompts; build semantic layers; then operational layers (actions, policies, events, skills); then outcome layers (state models, intents, KPIs, federated ontology and knowledge graphs as memory).
White-box reasoning graphs showed how a mother agent can orchestrate child agents, with pivot agents generating hypotheses in parallel, variable mapping between steps, and evals or guardrails at specific nodes — not only at the edge of the system.
Key points
- Three levels of agent maturity: assistive, automation, autonomous — and how engineering changes at each.
- Context engineering stack: semantic → operational → outcome layers.
- White-box reasoning graphs with node-level guardrails and memory.
- Event-driven autonomy via Kafka / business-object signals — no chat UI required.
- Live sales control tower demo with parallel hypotheses to cut token cost.
TakeawayPrototyping an agent is easy. Architecting one that runs in production — with governance, memory and observability — is the real engineering challenge. That is where the future of engineering is headed.