Agentic AI systems can look impressive in a controlled demo. A prompt goes in, the system reasons through the request, calls the right tools, and produces an answer or completes a task. But moving from that controlled environment to production introduces a very different set of challenges.
At QCon AI Boston 2026, Venugopal Jigidam our Director of Engineering, shared lessons from his experience building and scaling WaveMaker's Agentic Application Generation Platform. He explores how the platform evolved from an early pilot into a production system with multiple agents, more than 200 tools, and thousands of instructions supporting enterprise application development. WaveMaker's platform itself is designed around an architecture-first approach, with agents, enterprise guardrails, design systems, and a two-pass generation model.
His talk explores a central question facing teams building agentic AI systems today: Why do systems that work so well in demos often struggle when they reach production?
The answer, he argues, is not necessarily a problem with the underlying model. It is often a systems, architecture, and control problem.
Why Agentic AI Systems Work in Demos but Break in Production
A pilot typically operates within carefully controlled boundaries. There is limited context, a small number of tools, a narrow task, and relatively predictable inputs.
Production removes many of those artificial constraints. As agentic AI systems scale, they encounter larger knowledge bases, more tools, more complex workflows, application-specific context, and unexpected edge cases. During the talk, Venu walks through five challenges that emerged as the WaveMaker platform moved from experimentation toward production.
1. Context Overload Leaves Less Room for Reasoning
An LLM receives much more than the user's request. Its context can include system instructions, application information, chat history, documentation, workflows, metadata, and tool definitions.
As these inputs accumulate, the context becomes bloated. The model has to spend more effort filtering irrelevant information, leaving less room for reasoning. This can increase hallucinations, inconsistency, latency, and token costs.
The solution explored in the talk is progressive disclosure: give the model information when it needs it rather than loading everything upfront.
Skills make this possible by allowing agentic AI systems to expose capabilities on demand. Instead of sending every security instruction to an agent when the user only needs OAuth, the relevant skill can be loaded when required.
2. Tool Explosion Makes Agentic AI Systems Harder to Scale
Enterprise agentic AI systems can require hundreds of tools to interact with databases, APIs, security systems, repositories, and other enterprise services. But every tool definition consumes context.
Venu describes how WaveMaker's platform grew to more than 200 tools and encountered the resulting impact on context size and performance. The answer was not to remove capabilities, but to make tool access more selective through tool search.
Instead of loading every available tool, the system can retrieve the tools relevant to the task and expose only those to the model.
3. Agent Orchestration Can Create Information Loss
The next challenge emerged from agent orchestration. The initial architecture used different agents for different product capabilities, with an orchestrator delegating work between them. While this provided clear domain ownership, every handoff introduced another coordination step.
An agent might discover an important business rule while examining a database, but that insight could be lost when the task was summarized and handed to another agent. This is where the distinction between an agent, skill, and tool becomes important:
Agent: owns an end-to-end responsibility.
Skill: provides a capability within that responsibility.
Tool: performs the actual action.
For agentic AI systems, choosing the right abstraction can reduce unnecessary handoffs while preserving the context needed to complete a task.
4. Execution Becomes a Black Box Without Observability
A conventional application can usually be debugged through logs and distributed traces. Agentic AI systems are more complicated.
A single request can trigger multiple model interactions, tool calls, responses, retries, and changes in context. When execution starts drifting several steps into a workflow, looking only at the final output does not reveal what went wrong.
Venu emphasizes the importance of observing the entire execution: what context the model received, which tools it called, what those tools returned, and how the workflow progressed. Observability therefore needs to be treated as a foundational part of agentic AI platforms, not something added after deployment.
5. Agentic AI Architectures Cannot Remain Static
Perhaps the biggest lesson from the experience is that agentic AI systems require continuous architectural evolution.
The WaveMaker team changed its architecture four times in one year as new capabilities and patterns emerged. What worked at one stage did not necessarily remain the best approach as the underlying technology evolved.
The goal was not simply to make the model more capable. It was to make the overall system more predictable.
What These Challenges Reveal About Agentic AI Systems
The five challenges point to a broader lesson: building reliable agentic AI systems requires engineering discipline around the model.
Context needs to be managed rather than dumped into the model. Tools need to be selectively exposed. Agent boundaries need to be deliberate. Execution needs to be observable. And architecture needs to evolve alongside rapidly changing AI capabilities.
In other words, the model is only one part of the system. For enterprises building agentic AI solutions, the architecture surrounding that model ultimately determines how predictable, maintainable, and controllable the system can be in production.
And that is perhaps the biggest takeaway from Venu's experience: agentic AI systems are built through engineering, not just AI.
Watch the full talk to go deeper into the challenges of building production-ready agentic AI systems, see the examples behind these failure patterns, how WaveMaker approached them, and what the experience reveals about building reliable agentic applications at scale.
FAQs
1. What is agentic AI and how does it work?
Agentic AI enables software to interpret a goal, reason about the steps required, use tools, and execute tasks with varying levels of autonomy. Agentic AI systems typically combine an AI model with context, tools, skills, memory, and orchestration.
2. What features should I look for in an agentic AI orchestration platform?
An agent orchestration platform should provide context management, tool discovery, clear agent boundaries, skills-based execution, observability, tracing, guardrails, and enterprise integrations. These capabilities help agentic AI systems remain manageable as complexity grows.
3. What are the key benefits of using an enterprise agentic AI platform?
Enterprise agentic AI platforms can automate complex workflows, connect AI to enterprise data and systems, and help development teams scale AI-driven applications. For production environments, capabilities such as governance, observability, context control, and predictable execution are equally important.
