Building a Deterministic Software Factory for AI Agents

The Shift from Vibe Coding to Engineering

IndyDevDan argues that the current industry trend of 'vibe coding'—relying solely on LLM agents to perform tasks without oversight—is unsustainable for production-grade systems. The solution is a 'Software Factory' architecture that treats code as a first-class citizen alongside agents. By integrating deterministic code checks between agentic phases, engineers can reduce hallucinations, lower costs, and create repeatable, scalable development workflows.

The Three Pillars of the Software Factory

To move beyond basic prompt engineering, the factory is built on three non-negotiable principles:

1. Observability: Every phase, prompt, tool call, and cost breakdown must be logged in a swim-lane view. If you cannot measure the agent's output, you cannot improve the system.
2. Customizability: The system uses a unified YAML configuration to manage the 'core four' of agentic design: context, model, prompt, and tools. This allows developers to swap models (e.g., Gemini 3.6 Flash for cost-efficiency, Claude Opus 5 for complex reasoning) without changing the underlying workflow logic.
3. Reusability: The system is designed as a portable skill that can be installed into any codebase via a single command, allowing agents to orchestrate their own deployment and execution.

Harness Engineering and Deterministic Gates

Instead of relying on agents to verify their own work, the factory uses 'harness engineering.' This involves writing deterministic code to validate outputs (e.g., running linters, type checks, and test suites) between agentic steps. If a gate check fails, the system automatically routes the work back to the build agent for correction. This approach prevents the inefficiency of passing successful test results back into the LLM's context window, saving tokens and reducing error rates.

Workflow Orchestration

Dan demonstrates an AI Developer Workflow (ADW) that progresses from simple scouting to full Software Development Life Cycle (SDLC) execution. By defining clear phases—plan, build, test, review, and commit—the system creates a reliable pipeline. The factory uses typed JSON envelopes to pass context between agents, ensuring that each step has exactly the information it needs, rather than dumping entire codebases into every prompt.