Shopify's Roast is a convention-oriented workflow orchestration framework that structures AI into repeatable, reliable pipelines. Its CodingAgent integration with Claude Code brings agentic autonomy within defined guardrails — representing a practical middle ground between fully autonomous agents and rigid automation.
How It Works
Roast emerged from Shopify's Augmented Engineering DX team, which tackles developer productivity challenges like flaky tests and low test coverage using AI agents. They discovered that breaking complex tasks into discrete steps was key to reliable AI performance.
The Framework
Roast uses a declarative approach: YAML configuration files define workflows, and markdown files define prompts. This separation means:
- Non-AI engineers can understand and modify workflows without touching code
- Workflows are version-controlled and reviewable like any other config
- Steps are composable — mix deterministic operations with AI reasoning
Each workflow is a sequence of steps. Steps can be:
- Deterministic — Run linters, formatters, test suites, code transformations
- AI-powered — Use an LLM to reason about code, generate fixes, or make decisions
- Agentic — Hand off to the CodingAgent for autonomous iteration
The CodingAgent
The CodingAgent is what makes Roast distinctive. It's a full integration with Claude Code that brings agentic capabilities into structured workflows. You define the guardrails and objectives; the agent has autonomy within those boundaries to iterate, test, and improve until it achieves the goal.
Real Example: "Boba" (Sorbet Type Addition)
Shopify's "Boba" workflow for adding Sorbet type annotations to Ruby code demonstrates the pattern:
- Deterministic steps clean up code and run Sorbet's autocorrect
- CodingAgent takes over for remaining type errors — iteratively fixing issues, running tests, and ensuring everything passes
- The result is a typed file that passes CI, with the AI handling the ambiguous cases that autocorrect can't
Architectural Philosophy
Roast embodies a key insight shared with Stripe's Minions: structure the deterministic parts deterministically, and only use AI for the parts that require reasoning. Don't ask an LLM to run a linter — run the linter, then give the LLM the results.
The framework's convention-over-configuration approach means less boilerplate and more focus on the actual AI task. YAML + markdown is a deliberately low-barrier format that doesn't require deep framework knowledge.
Why It's in Assess
Roast is open source and usable today, which puts it ahead of most internal-only systems. The YAML + markdown approach is refreshingly accessible, and the CodingAgent integration with Claude Code is a compelling pattern for teams already using that tool. However, the framework is still relatively new and extracted from Shopify's specific needs (Ruby/Sorbet ecosystem). Assess it for your own workflow automation needs — particularly if you have repetitive code quality tasks that blend deterministic tooling with AI reasoning.
Key Characteristics
| Property | Value |
|---|---|
| Company | Shopify |
| System | Roast + CodingAgent |
| Architecture | YAML-configured workflow pipelines with agentic steps |
| Key innovation | Convention-oriented AI workflow orchestration with Claude Code integration |
| Best suited for | Repetitive code quality tasks (typing, test coverage, migrations) |
| Open source | Yes — shopify/roast |
| Language | Ruby |
| AI integration | Claude Code (CodingAgent) |
| Sources | Shopify Engineering: Introducing Roast |