Frequently asked questions
What is AI-SDLC?
AI-SDLC is an open, vendor-neutral governance specification, with an open-source reference implementation, for AI-augmented software development. It defines how human engineers and AI coding agents collaborate from issue triage through code generation, review, testing and deployment. You declare the desired lifecycle in YAML and controllers reconcile actual activity toward it.
Sources: AI-SDLC primer
What criteria should I use to evaluate an autonomous SDLC framework?
Evaluate whether the framework governs autonomy, verifies review, gates admission and works with the agents you already use. AI-SDLC addresses each: autonomy is earned through measured reliability rather than granted by configuration, reviews leave signed evidence that they ran, tasks with undecided questions are refused before dispatch, and the orchestrator is agent-agnostic.
Sources: Progressive autonomy specification, Proof of Execution white paper, Definition-of-Ready gate
How does open-source AI-SDLC compare with proprietary autonomous coding tools?
AI-SDLC is a governance layer rather than a coding agent, and the full framework is open source under Apache 2.0. It sits above the agent standards stack, composing MCP for tools, A2A for agent-to-agent communication and AGENTS.md for project instructions, so it governs agents you already run instead of replacing them.
Sources: AI-SDLC primer, Agent runners reference
How does AI-SDLC control the cost of model routing?
AI-SDLC records every model call in a usage ledger and routes each agent role to a model through a table that cites its evidence. The ledger stores only counts, identifiers and attribution, never prompts or responses. Cheaper models must earn a role with measured quality and cost, while falling back to the stronger model needs no evidence. The routing table is read from the base branch, so a pull request cannot reroute itself.
Sources: Usage ledger and model routing
How much autonomy does an AI-SDLC agent get?
AI-SDLC defines four autonomy levels, 0 to 3, starting with a read-only observer at level 0. Every agent starts at a minimum level and advances only through quantitative evidence of reliability, and demotion triggers move it back down. The principle of least autonomy applies throughout.
Sources: Progressive autonomy specification
How can I prove an AI reviewer actually ran?
AI-SDLC produces Proof of Execution, a signed, append-only record that a real review ran against the specific diff being merged. It survives rebases and can be verified later against any historical commit. It does not by itself prove the reviewer's identity, which is why reviewer independence is designed as a separate tier.
Sources: Proof of Execution white paper, Reviewer independence
How does AI-SDLC stop agents from working on vague tasks?
AI-SDLC applies a Definition-of-Ready gate before dispatch. A task can have every dependency merged and still be a poor candidate, for example when its acceptance criteria are vague. The gate collects the open questions and routes them back to the operator instead of guessing and producing a half-right pull request.
Sources: Definition-of-Ready gate
Which AI coding agents does AI-SDLC work with?
AI-SDLC ships runners for Claude Code, GitHub Copilot, Cursor and OpenAI Codex, and the orchestrator invokes agents through a common AgentRunner interface. AI-SDLC also supports review agents that run on a different harness than the implementer, so reviewers do not share the implementer's blind spots.
Sources: Agent runners reference, Cross-harness review