AWS's AI-DLC Workflow Runs 33 Stages Across 7 Coding Agents, and 230 Issues Are Open

AWS Labs' aidlc-workflows repository packages a methodology called AI-DLC as steering rules that seven coding agents can load. The README lists 5 phases, 33 stages, 14 agents and 11 workflow profiles, and the repository shows 230 open issues as of October 16, 2026.

The task it takes on is one most teams using Claude Code or Cursor have met: the agent writes code before anyone has agreed what it should write. AI-DLC puts a stop at each approval point and keeps a log of what happened. This explainer covers what the repository contains and what its tracker says about how well the stops hold.

What AI-DLC is

AI-DLC stands for AI-Driven Development Life Cycle. AWS's DevOps blog, in a post dated November 29, 2025 by Will Matos, Raj Jain, Siddhesh Jog and Raja SP, describes it as assigning routine tasks to AI while humans keep the critical decisions.

That post describes three phases. Inception covers planning and architecture, Construction covers design and implementation, and Operations covers deployment. The depth scales to the task: a simple bug fix skips planning, while a complex feature gets requirements analysis, architecture design and detailed testing.

The repository has grown past that. Its v2.11.0 README lists five phases and 33 stages, and developertoolkit.ai, which sells a paid toolkit of its own, reports the phases as Initialization, Ideation, Inception, Construction and Operation.

How it runs

A user installs an aidlc command, runs aidlc config --harness <name>, then invokes /aidlc or $aidlc in Codex CLI with a request. The workflow chooses a path, asks for missing decisions and halts at approval gates. A state file tracks progress, and an audit trail records prompts and responses. The README counts 119 audit events, while the third-party glossary cites 118.

Harness	Listed support
Claude Code	Yes
Kiro CLI and Kiro IDE	Yes
Codex CLI	Yes
Cursor, opencode	Yes
GitHub Copilot	Yes

The project is MIT-0 licensed, has about 5,300 to 5,400 GitHub stars, 959 forks and 1,592 commits. The README recommends Claude Opus 4.8 and says the workflow works best with capable reasoning models.

Stars measure interest and nothing about use, and the repository gives no adoption figure.

What the tracker shows

The 230 open issues mostly concern the gates themselves. On October 16, issue #2706 reported that some guard refusals leave no audit row, which undercuts the audit trail's purpose. Issue #2702 reported that a retried review with a malformed second file cannot be recorded.

Issue #2597 describes a product-lead review refused for a second Reviewer line, with the retry losing the first review's findings. Issue #2654 reports that --test-strategy is silently ignored.

Harness problems are the other cluster: #2582 covers GitHub Copilot CLI subagent calls under their own session id, and #2610 covers a Kiro IDE turn that can end with no next step named. Issue #2712 reports that aidlc config writes a Bedrock profile and region into tracked harness files.

None of the visible issues mention token cost, so the tracker says nothing about what a 33-stage run spends.

What the README warns about

The README says "Generative AI can make mistakes" and tells users to review output and costs before acting. It also warns that Kiro IDE's .kiro hooks can run commands on a machine, so only projects you own or have checked should be trusted.

The third-party glossary adds a warning of its own: that beyond throwaway work nothing replaces a human reading the code, and developers can drift into passive execution. Rituals without verification are a risk, it says.

Who should try it

If your team already uses one of the seven listed agents and wants approvals recorded, the audit trail is the feature to test first, and issue #2706 says to check whether every refusal is logged. If you want a light prompt file, 33 stages is more process than you asked for.

Watch the tracker rather than the star count: the repository reports 88 open pull requests, so the fixes for these gaps are already queued. For other agent-related open-source projects, see our reports on REA and DEsk.