Technical Tuesday: where you build your agents matters more than you think

Reasoning models are getting very good. Claude, GPT, Gemini can parse documents, write code, make judgment calls, handle ambiguity. Two years ago, this was science fiction. And yet, 95% of AI projects* still don't make it to production.

The biggest barrier to AI adoption is your own ability to contextualize and workflow-engineer these models.”

Jerry Liu, founder and CEO, LlamaIndex

The models aren't the problem. Everything around them is. If you're deciding where to build your agentic AI workflows, here's the question that should guide you: not which platform has the coolest demo, but which one will actually get you to production and keep you there.

Strong deterministic automation paired with AI

Teams that succeed in getting their AI projects to production understand that real-world agentic applications are a mix of deterministic flows with pockets of AI reasoning. For example, a travel approval workflow isn't pure AI. It's:

• A user submitting a request (deterministic)

• An agent extracting rates from messy policy documents (AI reasoning)

• A manager approving the trip (human-in-the-loop)

• Finance signing off (human-in-the-loop)

• Booking the travel (deterministic again)

The AI part is just a piece of the workflow. But if you can't orchestrate the remaining parts and observe the whole thing end to end, you don't have a production-ready system. You have a demo.

The majority of horizontal dev platforms give you great building blocks for the AI pieces. But that’s not enough. You also need enterprise-grade orchestration that stitches together AI agents, human approvals, deterministic logic, and unified observability across all of it.

A different approach to agentic workflows

UiPath comes to this space from a background rooted in enterprise automation. That perspective shapes how the UiPath Platform™ supports agentic workflows today.

Orchestration

UiPath Maestro™ orchestrates entire processes (people, deterministic automation, and agentic reasoning) in one flow. You see where a process is, who owns the next step, what the AI decided, and why. Most platforms make you stitch together separate tools for each of these. That stitching becomes technical debt. It slows you down and creates blind spots.

Observability

When something breaks in a multi-step agentic process, you need to know what happened. Not just in the large language model (LLM) calls but across the entire workflow. UiPath gives you end-to-end traceability. AI reasoning logs interleaved with deterministic process logs. You can inspect the prompts sent to models at every step while also seeing human approvals, system integrations, and business logic.

As Liu put it after demoing the integration: "No other platform, to my knowledge, actually does this—where you can get end-to-end traceability, observability, and governance across both LLM calls and deterministic process logs."

See Liu talking with UiPath CEO and co-founder Daniel Dines onstage at UiPath FUSION 2025. Register for The Best Bits and you'll get the recording of their full keynote delivered right to your inbox in Volume 1.

Governance and the AI Trust Layer

AI agents making decisions in production need guardrails. UiPath AI Trust Layer gives you centralized control over every generative AI interaction across the platform. Personally identifiable information (PII) masking ensures sensitive data is pseudonymized before it ever reaches an LLM. You get usage auditing, cost controls, and a secure LLM gateway that enforces your data governance policies. Bring your own LLM if you need AI sovereignty or cost control—your models still inherit all the platform's guardrails. This isn't security bolted on as an afterthought. It's native.

*MIT NANDA, The GenAI Divide STATE OF AI IN BUSINESS 2025, 2025.