Reference
Prerequisites
What you need before adopting the AEF framework. Three categories of requirements and a list of nice-to-haves.
Technical requirements
The baseline infrastructure your project needs. These are hard requirements; without them, the workflow cannot function.
Git-based workflow
your codebase must be managed with Git. The workflow uses branches, worktrees, and pull requests as its primary coordination mechanism.
CI/CD pipeline
any provider works (GitHub Actions, GitLab CI, Jenkins, CircleCI). The framework needs a way to run automated checks on commits and pull requests.
Test suite with >60% coverage
the Test phase and its quality gates depend on existing tests. Below 60% coverage, the feedback loops lack enough signal to self-heal effectively.
AI tooling
The AI infrastructure that powers the workflow agents.
LLM with tool-use capability
Claude, GPT, or equivalent. The model must support tool use (function calling) so agents can read files, run commands, and interact with your codebase.
API key or CLI tool configured
either direct API access or a CLI tool like Cline or Kiro. The runner abstraction supports both patterns.
Team knowledge
The human skills needed to build and maintain the agentic layer.
Familiarity with prompt engineering basics
you don't need to be an expert. Understanding how to write clear instructions, provide context, and structure outputs is sufficient to start.
Understanding of your codebase's architecture
the agentic layer encodes your project's conventions and structure. Someone on the team needs to know how the pieces fit together.
Comfort with code review processes
at Level 2-3, humans still review agent output. Your team should have an established code review culture, even if informal.
Recommended (not required)
These accelerate adoption but are not blockers. You can build them as you go.
AGENTS.md or similar
an existing agent instructions file gives your agentic layer a head start. It already encodes project context that prompt templates need.
Documented coding standards
style guides, naming conventions, and architecture decisions. These become the basis for your review phase quality gates.
Automated linting/formatting
tools like ESLint, Prettier, Black, or Ruff reduce noise in the review phase. Agents produce cleaner output when formatting is automated.