follow us on x:
https://x.com/vaibcode
https://x.com/dexhorthy

The AI coding stack is becoming modular, but deciding what to build, what to buy, and how to connect it all is becoming its own engineering problem. In this episode of AI That Works, Dex and Vaibhav break down the architecture behind modern software factories and explore how teams can build systems that let AI agents write, test, review, and ship software with less human intervention.

The conversation starts with the biggest decision facing teams building agent infrastructure: do you build the entire stack yourself, or buy a fully managed system? From there, they map out the layers underneath a software factory, including development environments, coding harnesses, and the control plane. They explain how teams can mix and match these pieces, why the development environment becomes increasingly important as systems get more complex, and how enterprise teams can maintain the flexibility to bring their own infrastructure.

They also walk through real examples of agent workflows, including headless Claude Code and Codex sessions, automated CodeRabbit review loops, agent-driven merge queues, and orchestration systems that connect Slack, Linear, GitHub, and other tools. Rather than trying to automate everything immediately, they argue for building toward systems that are roughly 95% automated while keeping humans involved where judgment or external information is still required.

The episode ultimately makes the case for composability. As AI coding agents continue to evolve, teams should be able to swap models, harnesses, compute, and orchestration layers without rebuilding the entire system. Dex and Vaibhav explore the emerging interfaces between these layers, why today's agent ecosystem still lacks strong standards, and how the future may look less like one giant AI platform and more like an open software stack where engineers can choose exactly what they want to build, buy, and control.

KEY TAKEAWAYS
• Software factories can be built from interchangeable layers rather than one monolithic platform
• Teams need to decide which parts of the AI coding stack are worth building and which are better purchased
• The development environment becomes a major engineering challenge once agents need access to complex application stacks
• Automated review loops can push human intervention later in the development process
• Most teams should aim for highly automated workflows rather than trying to eliminate humans entirely
• Control planes connect agents to the tools, workflows, permissions, and information they need
• Enterprise teams often need to own their compute and development environments while keeping the agent layer flexible
• Multiple repositories can be coordinated through a shared workspace without forcing everything into a monorepo
• Standard interfaces between AI harnesses and orchestration systems are still immature

TIMESTAMPS
00:00 Build vs. Buy: The Future of Software Factories
00:24 What This Episode Covers
01:31 What Makes a Software Factory Actually Work
02:47 Why Codex Is Taking Over AI Coding Workflows
05:33 Why AI Coding Workflows Are Becoming Personal
08:49 The Hidden Cost of Switching AI Models
11:08 Building a Workflow for Large AI Coding Projects
15:32 Automating the Feedback Loop From User Reports
19:34 Building Evals for Automated Issue Triage
22:08 Why Software Factories Need Multiple Layers
25:40 Building an Agent-Powered Merge Queue
27:16 How Agent Work Gets Dispatched to the Right Machine
29:41 What Makes an Agent Workflow Reliable?
33:47 Why the Development Environment Matters So Much
36:23 Building the Inner and Outer Agent Harness
39:55 Why the Control Plane Is the Most Underserved Layer
42:19 Pets vs. Cattle: How to Manage Agent Environments
45:30 How to Manage AI Work Across Multiple Repositories
50:00 Where Software Factory Design Patterns Are Going
51:23 Why AI Infrastructure Needs Better Interfaces
54:16 The Interfaces Connecting Agents to Infrastructure
55:56 Why You Shouldn't Build the Entire AI Stack Yourself
1:00:12 Why AI Infrastructure Needs to Be Composable
1:05:20 Why Companies Want to Own the AI Harness
1:05:40 When Software Becomes Its Own Extension Platform

TOPICS COVERED
• Software factories
• AI coding agents
• AI engineering
• Agent orchestration
• AI coding harnesses
• Claude Code
• Codex
• AI infrastructure
• Agent workflows
• Autonomous coding
• CodeRabbit
• Agent merge queues
• Control planes
• Development environments
• Cloud development environments
• Compute infrastructure
• AI evals
• Human-in-the-loop AI
• Compounding engineering
• MCP
• ACP
• AG-UI
• Multi-repository development
• Git submodules
• Git subtrees
• Composable AI infrastructure
• Enterprise AI
• Open source AI infrastructure

HASHTAGS
#AIThatWorks #AICoding #AIEngineering #CodingAgents #SoftwareFactories #ClaudeCode #Codex #AIAgents #SoftwareEngineering #AgenticAI #AIInfrastructure #DeveloperTools #MCP #ArtificialIntelligence