Zhou Yu, co-founder of Arklex AI and a professor at Columbia University, has a blunt assessment of the AI agent landscape: 95 percent of agents remain stuck in demo mode. They are impressive in controlled settings, but they cannot reliably perform the real-world tasks that would justify deploying them in production. At a recent InfoQ presentation, he laid out why the gap exists and what his team is building to close it.
Why demos do not survive contact with compliance
The demo-to-production gap is not a technical limitation in model capability. It is a failure of evaluation. Zhou walked through a voice agent demo for a credit card application, a conversation that covered product recommendation, personalized rewards explanation, application submission, and post-approval follow-up. The demo works. The agent handles multi-turn dialogue, integrates with backend systems, and completes a complex transaction. But it is not in production.
The reason is compliance. In finance, healthcare, and other regulated industries, deploying an AI agent requires demonstrating to a compliance team that the system behaves correctly across every possible interaction. The traditional approach to model validation relies on static benchmarks: a set of question-answer pairs with known correct outputs. Run the model, compare its answers to the reference answers, report a score. This works for single-turn systems where each query is independent.
Multi-turn agents break that model entirely. A user's follow-up question depends on what the agent said in the previous turn. If the agent changes its behavior, the entire conversation trajectory shifts. Tool calls modify external state: account balances change, products get returned, applications get submitted. You cannot evaluate the agent by looking at its text output alone. You have to look at what happened in the database, whether the user's actual intention was fulfilled, and whether the action was performed correctly.
The manual testing bottleneck
Today, most teams evaluate multi-turn agents by doing it by hand. You build a chat agent, publish it as a web service, and ask your friends, coworkers, or product manager to try it. They chat with the agent, find errors, and report them back. You cluster the errors, identify root causes, fix the agent, and deploy again.
Then you start over. The old test data is invalid because the agent changed. You ask your friends to test again. The process is slow, labor-intensive, and produces poor coverage. The testers are not your real users. They are technical people testing ideas, not real customers with real needs. When you finally deploy, the users you actually get are talking about things you never imagined from a developer's perspective.
Zhou describes this as the core problem his team set out to solve. The testing loop is too slow, too dependent on human labor, and too far removed from actual user behavior to produce reliable results.
Synthetic users as the testing backbone
The solution Arklex is building uses AI agents to simulate users. Instead of asking your friends to test your shopping agent, you spin up a collection of synthetic user agents, each with different personas, goals, and behavioral patterns. These simulated users interact with your production agent through the same API endpoints that real users would hit.
The approach produces trajectories that approximate how real users will behave, but at a fraction of the cost and with full control over coverage. You can constrain specific user agents to test particular components or functionalities. You can run dozens of them in parallel. And because the synthetic users are API endpoints, the entire process integrates into a CI pipeline.
The open-source tool for this is called ArkSim. It takes three inputs: the agent's capabilities and knowledge, a set of scenarios defining what the simulated users should try to accomplish, and configuration for how many conversations to generate per persona. Once configured, the pipeline runs automatically on every code push, producing evaluation data without human involvement.
What the simulation output actually looks like
In Zhou's demo, the setup starts with a description of the agent's capabilities and a knowledge file containing business context and frequently asked questions. The scenarios define simulated users with different education levels, purchase histories, and interaction goals. ArkSim generates conversations between these users and the agent, then evaluates the outcomes against the defined scenarios.
The key difference from manual testing is repeatability. Every code change triggers the same set of scenarios, so you can measure whether a fix for one error introduced regressions elsewhere. The synthetic users do not get tired, do not have scheduling conflicts, and do not report errors in vague terms that require follow-up clarification.
Where this fits in the broader agent ecosystem
Zhou's framing of the problem aligns with what production teams are discovering. Shopping agents like Walmart's Sparky and Amazon's Rufus have been in production for over a year, operating in the relatively low-risk domain of product recommendations and purchase assistance. Financial and healthcare agents, where the consequences of errors are higher and the regulatory burden is heavier, remain largely in demo or limited-release status.
The gap between these two categories is not model capability. It is the ability to demonstrate, with evidence, that the agent will behave correctly across the full range of user interactions it will encounter. Synthetic user simulation is one approach to generating that evidence at scale. Whether it satisfies compliance teams is the next question, and one that the industry is still working through.