Reliability engineering for enterprise AI agents: evaluations, simulations and episodic memory

Описание: Shipping AI agents in the enterprise requires more than a convincing demo. https://uipath.ly/46f0drb

Featuring UiPath AI Leaders:
Taqi Jaffri, VP, Product Management
Scott Florentino, Director, Software Engineering

In this episode, we break down the core pillars of reliability engineering for enterprise AI agents: evaluations, simulations, and episodic memory. Taqi and Scott walk through the practical systems that move agents from prototype to production.

You’ll learn:
What agent evaluations actually measure (context, precision, faithfulness)
How deterministic evaluators compare to LLM-as-a-judge approaches
Why trajectory evaluation matters beyond simple input/output testing
How to simulate tool calls, escalations and failure conditions
What episodic memory is — and how it differs from fine-tuning
How agents improve over time using structured human feedback
How to measure reliability before release

If you're building AI agents that need to be auditable, measurable and production-ready, this episode focuses on the engineering discipline behind enterprise-grade deployment.

(This video was recorded in November 2025)

⏱ Timestamps

00:00 Why reliability engineering matters for enterprise AI agents
00:24 What are agent evaluations?
01:14 Context, precision and faithfulness explained
02:07 Evaluation sets and LLM-as-a-judge
03:17 Generating strong evaluation data
04:24 Trajectory evaluation vs black-box testing
05:30 Simulating tool calls and failure scenarios
07:08 What is episodic memory?
08:15 Episodic memory vs fine-tuning
09:14 Learning from human feedback
10:15 Managing and inspecting agent memory
11:30 Defense-in-depth for enterprise AI reliability
12:15 Proving performance with measurable reliability scores

🚀Join our community for more updates:
Academy: https://uipath.ly/462jPft
Blog: https://uipath.ly/3EwSRRC
LinkedIn: https://uipath.ly/44SelTOa
Facebook: https://uipath.ly/45A4naF
Forum: https://uipath.ly/3ReA7O3