Applied Intelligence
Module 1: Introduction to ASD

The Developer's New Role

Architect, director, reviewer: the mental model shift

The shift from traditional development to ASD changes what a developer does, not whether a developer matters.

In traditional workflows, developers occupy a single role: implementer. The developer reads requirements, designs solutions, types code, runs tests, debugs failures, and iterates until the code works. Every keystroke flows through human hands.

ASD introduces role separation. The developer now operates as architect, director, and reviewer sometimes within the same hour.

The three roles

Architect Define system boundaries before any code is written. Which patterns to use, how components interact, where to draw module boundaries these decisions remain with the developer.

Director Orchestrate agent work rather than performing it. Deploy agents to different tasks, monitor progress, redirect when approaches fail. The skill is problem decomposition.

Reviewer Validate every line before it reaches production. Research indicates only ~30% of AI-suggested code gets accepted as-is. Verification is the core activity.

Architect: Defining boundaries

Agents excel at implementation but lack the context to make architectural decisions. An agent asked to "add authentication" will produce working code, but whether that code uses JWT tokens versus session cookies versus OAuth depends entirely on the constraints the developer establishes.

Director: Orchestrating work

Boris Cherny, creator of Claude Code at Anthropic, describes running five to ten agent sessions simultaneously. The workflow resembles real-time strategy games more than typing: deploying agents to different tasks, monitoring progress, redirecting when approaches fail.

The skill is not typing speed but problem decomposition breaking work into chunks an agent can execute independently.

Reviewer: Validating output

Research indicates only about 30% of AI-suggested code gets accepted as-is. The remaining 70% requires modification or rejection.

Developers report spending more time reviewing agent output than they previously spent writing equivalent code manually and finding that distribution correct.

The mental model shift is fundamental: from producing code to producing outcomes through code.

On this page