Agent-First Development

What is Agent-First Development?

Agent-first development is a workflow where the developer starts by instructing an AI agent to write code, then reviews and refines the output — rather than writing code first and using AI as an assistant. It represents a fundamental inversion of the developer’s role from primary author to technical director.

The Workflow Inversion

Traditional workflow:

1. Developer opens editor, writes code
2. Gets stuck or needs help → asks AI
3. AI provides suggestions → developer integrates

Agent-first workflow:

1. Developer describes what they want to the agent
2. Agent writes the code (with tool access, bash, file system)
3. Developer reviews diff, merges or redirects

As DHH described it: the shift went from “I’m code first, everything” to “Now I start with the agent.”

Why It Works for Senior Developers

Agent-first development is most effective for experienced developers who can:

• Validate output quality against production standards
• Spot architectural problems before they compound
• Redirect agents efficiently with domain-specific guidance
• Run multiple agents in parallel on related tasks

The key insight is that taste, judgment, and system knowledge become the constraint — not typing speed or implementation knowledge.

Practical Patterns

• Parallel model racing: Run two models simultaneously (e.g., Gemini for fast iterations, Opus for complex reasoning) and merge the best outputs
• Review-first commits: Check diffs in a Git UI before anything lands, treating agent output like a junior developer’s PR
• Exploratory prototyping: Use agents to explore hunches and ideas with near-zero investment — delete freely if the direction is wrong

Related Reading

• Agent Harness — The infrastructure that enables agent-first workflows
• Agentic Coding — The broader paradigm