AI agents in production: The architecture that works
Our second AI Build Week showcased what’s possible when you combine natural language prompting with enterprise-grade AppGen. We turned government web traffic data into live dashboards, converted spreadsheets into working apps, built churn prediction tools with Stripe, and vibe-coded production apps straight from lakehouse data.
Watch the recorded sessions:
Check out our YouTube channel for more demos and build-alongs.
AI agents in production: Why most teams fail (and how to fix it)
You don’t need total AI autonomy to create value. Agentic work is a spectrum, not a zero-sum game.
The teams succeeding right now are combining agentic workflows, AI agents, and human tasks for incremental gains with less risk.
The architecture that works
Most orgs actually don’t need to build one super-agent that does everything. Here’s the pattern we’re seeing succeed:
Orchestrator agents act as supervisors. Instead of one agent handling disputes end-to-end, create an orchestrator that delegates to specialized sub-agents. One analyzes language, another reviews customer history, another checks product usage—all running in parallel with clear guardrails.
Branching logic makes smart decisions. Based on what agents discover, the system routes automatically when evidence is clear-cut, or escalates to humans when there's ambiguity.
Human checkpoints are built in from the start. Humans review what AI discovered and make final calls—not as an afterthought, but as the safety net guiding and approving AI actions.
For example: You have a customer who disputes a charge. The orchestrator spawns multiple agents checking evidence, account signals, and usage patterns. If data is clear, the agent auto-resolves. If it’s ambiguous, the agent routes to a human with full context for approval.
Why chat interfaces aren’t the answer
ChatGPT went mainstream via chat, but prompting is exhausting for production workflows. You have to think of everything upfront, structure requests perfectly, switch between tools to enact different parts, and hope the agent parses correctly.
What works better: Purpose-built interfaces with structured inputs, file upload requirements, and email triggers for async processes. The UI acts as a control surface—letting people provide structured inputs and trace what AI is doing.
How to get agentic workflows into production
- Start with clear outcomes and explicit policies. Don’t ask AI to automate what you don’t understand yourself. Pick workflows with clear definitions of success.
- Invest in guardrails and observability. LLMs are non-deterministic. Track token usage, runtime, tool access, and security measures. Build automated evals to ensure output stays consistent.
- Add autonomy incrementally. One of our engineers built a tool to parse demo recordings. Rather than spending weeks integrating Zoom’s API, he started with a text box for manual transcript pasting. Two minutes per week of manual work, but creating value immediately while iterating toward full automation.
Watch the talk: AI Agents in Production from Retool Summit 2025
Real example: Komatsu’s agentic control panel
Eric Cheng, Enterprise Architect at Komatsu, built a Retool app that acts as a surface for building agentic systems to automate lengthy customer service operations. What started as a side project scaled easily—with Retool as a central control panel connecting external endpoints, file storage, APIs, and retail workflows.
The teams winning right now are combining orchestrated agents, deterministic processes, and structured human judgment into durable production systems.
Bonus content!
- Our CEO David Hsu was on the Neuron podcast talking about AppGen and Agents.
- Agents were mentioned in the Growth Unhinged newsletter this week.
- We dropped a new YouTube short on transforming a spreadsheet into an app.
- Is your pre-holiday calendar getting crazy? Check out this new guide to building a calendar concierge agent.
This is our second LinkedIn edition of this newsletter—what do you want to see? Let us know in the comments.