UiPath engineering blog

UiPath demonstrates a two-tier triage agent that uses Jev’s typed System One model for fast classification, confidence scoring, and escalation decisions before involving a human or full LLM. Live API measurements show Jev reducing routing latency by up to 23x at p50 and 32x at p99 while lowering costs.

Bai Li explains why measuring skill activation across coding agents requires agent-agnostic trace detection, multi-skill scoring, and awareness of truncated descriptions. The post presents a Coder Eval approach and reports that front-loaded explicit triggers increased micro-recall from 46.3% to 67.3% while maintaining approximately 0.96 precision.

UiPath compares Claude Haiku 4.5 and Sonnet 4.6 across more than 400 enterprise automation tasks, finding that similar short-task performance diverges sharply as workflows grow longer. The analysis explains error compounding, hidden grading differences, and how to reproduce the evaluation with the open-source Coder Eval harness.

This post compares three ways to restore state in long-running human-in-the-loop agent workflows: snapshotting, deterministic replay, and checkpointing. It uses a refund-processing example to explain determinism, idempotence, and the trade-offs each approach introduces for reliability and developer complexity.

This post explains how UiPath set up Temporal multi-cluster replication using mTLS for secure cluster-to-cluster communication over the public internet. It walks through the certificate lifecycle and Kubernetes-based automation with cert-manager, External Secrets, and Azure Key Vault to simplify CA sharing and server certificate issuance across primary and secondary clusters.

This post describes how UiPath built an AI-powered Case Classification Agent to infer support ticket metadata such as Deployment Type, Product, and Issue Type from freeform issue descriptions. It walks through several iterations—from retrieval-based approaches to BERT and finally RAG with ontology guardrails—to improve ticket routing accuracy and reliability.

This article explains how UiPath added a reliability layer on top of SignalR to support bidirectional communication between Apps runtime and robots. It covers message chunking, acknowledgments, retries, duplicate detection, and recovery strategies for large payload delivery.

This post explains how UiPath standardized contract-based end-to-end testing to make its platform shippable across cloud and on-premises environments. It outlines an ephemeral, hermetic test environment (ETE), the Athena execution contract, shared API clients, and declarative data provisioning to reduce flakiness, duplication, and environment drift. The article also describes how multiple test frameworks fit into the model and highlights the operational lessons learned from rolling out the approach at scale.

Delivering on the promise of real-time agentic automation requires a fast, reliable, and scalable data foundation, and this post explains how UiPath unified batch and streaming on Azure Databricks to support that need. It details the move to Spark Structured Streaming, the architecture changes behind lower latency and better scalability, and the trade-offs around delivery guarantees and operational simplicity.

UiPath explains how it built a secure natural-language search system that converts user questions into SQL without exposing unauthorized tenant data. The article covers prompt injection risks, column tampering, and a defense-in-depth design using inline table-valued functions and restricted SQL permissions.

UiPath describes moving telemetry from Azure Application Insights to OpenTelemetry plus Azure Data Explorer to reduce observability costs while keeping dashboards and alerts compatible. The new setup uses the OpenTelemetry Collector, ADX, and Grafana to lower telemetry spend and preserve developer visibility.

Smart Search is UiPath’s generative AI documentation assistant, using RAG, vector search, and GPT-4 to answer support questions from docs and knowledge base articles. It emphasizes version-aware filtering, daily reindexing, and fast support deflection across UiPath products.

This post explains how UiPath designed API Workflows as a lightweight engine for system-to-system automation at enterprise scale. It covers the platform’s core design principles, including security isolation, zero-copy execution, observability, and serverless runtime architecture, with a focus on handling user-supplied JavaScript safely and efficiently.