Why UiPath is re-designing its platform around agents that build automations, not just run them

Most enterprise AI conversations eventually hit the same wall. The agent did something unexpected. The process broke somewhere between systems. Nobody can tell you exactly where, or why, or how to stop it happening again. That is the gap UiPath is trying to close – and at a virtual product strategy session this week, the company made its clearest argument yet that solving it requires rethinking not just what automation platforms do, but who, or what, builds the automations in the first place.

But the bigger story is this: UiPath has pivoted its entire engineering roadmap to make its platform usable primarily by coding agents. That is a more substantive claim than it might first appear. The conventional framing of AI-versus-Robotic Process Automation (RPA) tends to focus on whether agents can replace automations at the execution layer. 

Co-founder and CEO Daniel Dines addressed this directly, drawing a line between what AI is good at and what a structured automation platform does well. He argues that AI is well-suited to understanding context and deciding which action to take, but for complex, multi-step processes running at scale – unattended and reliably – the infrastructure underneath still needs to be deterministic, governed, and auditable. That consistency is something probabilistic AI models are not designed to guarantee, and deterministic automation is. Dines observed:

I now see that really coding agents, it's an amazing accelerator.

What happens when coding agents are used at design time, not execution time, is where it gets more interesting. Chief Product and Technology Officer Raghu Malpani – who was appointed to the expanded role in late March, having joined UiPath as Chief Technology Officer in 2024 – laid out the ambition of coding agents handling the full automation lifecycle: 

They author and deploy, meaning they go from natural language to production-ready agentic workflows with guardrails.

From there, the same agents diagnose failures, propose fixes, and close the loop automatically.

In practice, this expands UiPath's addressable developer base significantly. The platform has historically targeted automation developers – a relatively specialist skillset. With coding agent support embedded the stack, professional developers who have never built a UiPath automation become viable users, and less technical users gain the ability to verify their intent against what the automation is actually doing through the low-code visual layer. Malpani made the case that low-code becomes more important in this model, not less – it gives people confidence that what they asked for is what got built.

Process orchestration vs. agent orchestration – why the difference matters

Maestro underpins all of this, and UiPath spent considerable time explaining why its implementation of process orchestration is substantively different from the agent-to-agent coordination that gets most of the industry attention. Dines made the strongest distinction: when enterprises talk about orchestration internally, they typically mean process orchestration – structured workflows that map compliance requirements, regulatory constraints, and the various human and automated actors involved in getting work done. That is a different problem from giving a swarm of agents a goal and letting them negotiate. Enterprise processes have hundreds of sub-workflows, many of which are deterministic, many of which require human approval, and all of which need to be auditable end-to-end.

One New Zealand, one of the country's three major telecommunications providers, offered the most credible illustration of this in practice. CEO Jason Paris described the challenge of a legacy business grown through acquisitions, running multiple billing platforms and systems that were never designed to integrate. The specific example he gave was a B2B handset replacement process that used to take four to five days end-to-end:

 We've changed that four-to-five days to five-to-10 minutes.

No re-platforming was required. Existing systems stayed in place, and Maestro provided the orchestration layer on top. Proof of concept to production took five weeks. Paris was straightforward about why that speed was possible – prior investment in RPA across the organization meant the data and processes were already understood, and UiPath's experts worked alongside One New Zealand's own team throughout. Paris was unambiguous about how far the transformation extends, noting: 

There is genuinely no part of our organization that is not going to be transformed through this technology.

It is a distinction that resonates with customers operating complex legacy estates, Paris continued: 

It avoids multi-tool complexity. It's an integrated tool that works in combination, not just as an orchestration layer, but across AI and RPA.

Vertical solutions – outcomes, not infrastructure

Newer, and representing a more significant market shift, is UiPath's vertical solutions push. Rather than selling automation infrastructure to IT, the ambition is to sell defined outcomes to business units – which requires the platform to come pre-configured with domain logic, industry-specific agents, and return on investment dashboards that speak the language of the business rather than the center of excellence. Mark Rubinstein, Director of Product Management leading the financial services vertical, walked through a loan origination quality assurance solution that demonstrates how this works in practice.

Mortgage lending QA is labor-intensive and error-prone – roughly 47% of critical defects in loans stem directly from manual verification and calculation, and the average cost to originate a single conventional mortgage is close to $11,000 – two-thirds of which is labor. Rubinstein's example solution aggregated data across loan origination, core banking, and content management systems that were not designed to talk to each other, runs automated checks against business rules, flags exceptions for human review, and generates audit-ready reports automatically. What used to take hours takes minutes. Rubinstein noted the solution was co-designed with regional banks and credit unions running it in production – this is not a reference architecture.

My take

Coding agents at design time, deterministic infrastructure at execution time – if that model works in practice, UiPath's total addressable developer base genuinely does expand, and the time-to-value argument strengthens considerably. The One New Zealand numbers are specific and honest enough to be credible. The loan origination solution addresses a problem that is real, measurable, and not going away. None of which addresses the execution pace – agentic case management launches in May, and native coding agent integration across the full lifecycle remains a roadmap commitment. There is still a distance between this investor session and those capabilities in customer hands at scale.