AI-DLC Workflows
One harness-neutral core, native across major AI coding agents, turning AI agents into verifiable, self-correcting engineering workflows.
Evidence shows approval gates (user confirmation) and audit logs, but least privilege, data flow transparency, sensitive data handling, dependency security, external effects, rollback, and source attribution are not adequately addressed. Deductions: no least privilege details, no data flow or sensitive data handling, dependency security not mentioned, external effects unclear, rollback not described, source attribution relies on unverified publisher.
Self-consistency is good: README, CI, and test files are consistent, but dependency availability is unverified (bun install relies on external network), and failure messages are not well demonstrated. Deductions: dependency availability unverified, failure messages not explicit in docs.
Audience and scenarios are clear (multiple harnesses), capability boundaries are described (stages, agents, scopes), trigger precision is high (commands and hooks), environment fit is good (multi-platform). Deductions: capability boundaries not fully explicit, environment fit relies on external tools.
Information architecture is clear (README, docs, dist), install notes are detailed, naming stability is present (version numbers), examples and FAQ exist, known limitations are mentioned, license is clear, versioning changelog exists but not detailed, maintenance responsibility not explicit. Deductions: naming stability not fully guaranteed, changelog not detailed, maintenance responsibility not explicit.
Output usability is high (generated workflows), marginal value is high (multi-harness), cost-benefit not fully assessed. Deductions: cost-benefit data not provided.
Claims are traceable (references to spec, blog), cross-source corroboration is limited (only internal docs), fact-inference separation unclear. Deductions: cross-source corroboration insufficient, fact-inference separation unclear.
- Publisher identity is unverified; treat source with caution.
- Dependency on bun installation requires secure network environment.
- AI-generated content requires human review to avoid errors.
What does this agent do, and when should you use it?
AI-DLC Workflows is an open-source implementation by AWS Labs that brings the AI-Driven Development Life Cycle methodology to life as executable engineering workflows. It defines a structured framework with 5 phases, 33 stages, a roster of 14 agents (11 domain experts, 2 reviewers, and an adaptive composer), and a 91-event audit trail, all contained in a harness-neutral core/ directory. Thin per-harness surfaces adapt this core to run natively on Claude Code, Kiro IDE, Kiro CLI, Codex CLI, Cursor, opencode, and GitHub Copilot. Users copy the generated dist/ files into their projects to enable the workflow. Every stage includes an approval gate, and the system learns from human corrections by turning them into persistent behavioral rules. The default scope is classic, with 11 adaptive scopes, 3 depth levels, and 3 test strategy levels, plus CLI tools to adjust workflow on the fly.
It loads the methodology definitions (stages, agents, rules, knowledge) from core/, adapts them to specific coding tools via harness layers, and executes a state-machine-driven workflow. Users start by invoking /aidlc in their coding session, optionally running /aidlc --doctor first to verify the setup. The engine orchestrates 14 agents (including domain experts and reviewers) across stages, pausing at each stage for human approval. It records 91 event types for auditability, supports session resume, jump-to-stage, and redo, and allows real-time changes to scope, depth, and test strategy. The deterministic engine is byte-identical across harnesses. Users can regenerate all harness distributions with bun scripts/package.ts.
- Development teams needing a structured, auditable AI-driven software lifecycle, who can install on Claude Code and run /aidlc to guide their project from ideation to operation.
- Organizations using multiple coding tools (e.g., Claude Code, Cursor, Codex) that want a consistent methodology and workflow across all agents.
- Enterprises requiring strict quality gates and traceability, thanks to built-in reviewer agents and 91-event audit logs.
- Teams that want to capture human corrections as persistent rules, improving agent behavior over time via the built-in learning loop.
- Developers who need to switch between different coding environments (local IDE, CLI, CI) while preserving the same engineering process.
What are this agent's strengths and limitations?
- Single methodology source generates all harness distributions, ensuring consistency and easy maintenance.
- Built-in 91-event audit trail and approval gates provide enterprise-grade traceability.
- Supports multiple major coding tools (Claude Code, Codex, Cursor, etc.), adapting to existing team workflows.
- Learning loop converts human corrections into persistent behavioral rules, continuously improving agent performance.
- Requires bun and proper PATH configuration for non-interactive shells, which can be tricky on Windows.
- Default configuration uses AWS Bedrock for Claude Code and Codex, requiring AWS account and model access, potentially adding cloud costs.
- Recommended model is Claude Opus 4.8; weaker models may degrade performance and may need paid subscriptions (e.g., Kiro plan).
- The process is structured with many approval gates, which might be cumbersome for small, rapid prototypes.
How do you install or deploy this agent?
- Install bun (common runtime): curl -fsSL https://bun.sh/install | bash (macOS/Linux) or irm bun.sh/install.ps1 | iex (Windows PowerShell). 2. Clone the main branch: git clone --branch main https://github.com/awslabs/aidlc-workflows.git. 3. Copy the harness distribution into your project, e.g., for Claude Code: cp -r dist/claude/.claude/ your-project/.claude/ and cp -r dist/claude/aidlc/ your-project/aidlc/, then merge AGENTS.md and .gitignore.
How do you use this agent?
For Claude Code: start claude in your project, run /aidlc --doctor to verify, then /aidlc Build a task management API with user authentication to begin. Follow the approval gates at each stage. Use /aidlc commands to check status, jump to a stage, or adjust scope/depth mid-flow.
How does this agent compare with similar options?
Compared to ad-hoc AI coding, AI-DLC introduces structured gates, role specialization, and knowledge retention, addressing context drift and lack of decision records. However, within coding-agent frameworks, alternatives like GitHub Copilot workspace or other workflow tools may exist, but no direct competitor is named in the source.