Req2Prod turns a plain-language requirement into shipped, production-grade software. Specialised CrewAI agents—orchestrated by a GitHub Actions pipeline, one agent per step—take a change from requirement and build, through automated code review, to deployment and post-deploy testing on the live service. The single human touchpoint is clarification: the Product Manager, Architect and Engineer can loop back to whoever wrote the requirement. Code review loops with a fix agent for up to three rounds, then escalates to an independent arbiter; once deployed, a production test can trigger an automatic rollback or a DevOps agent that fixes and re-deploys. The build front-end and the automated pipeline are joined by a single config switch—kept off today to control model spend—so the full requirement-to-production run is available on demand.
The agent crew & feedback loops
Delivery flow (build → review → deploy)
Clarification → requirement inputter (only human step)
Code-review fix loop (max 3 rounds) → pr_arbiter
Prod-test failure → rollback_agent / devops_agent
Production / live
◆ Requirement in, software out
Collapses the whole SDLC—analysis, build, test, review, release—into one autonomous, auditable flow.
↻ Self-correcting by design
Code review loops with a fix agent up to 3 rounds, then an independent arbiter decides; failed prod tests auto-rollback or invoke DevOps—no silent bad output.
⚙ Model-tiered & controlled
Each agent runs its own Claude tier (Haiku→Opus) from a model registry; the sole human step is requirement clarification. Tuned on cost, safe on risk.
req2prod
HOW IT'S BUILT · GITHUB ACTIONS + CREWAI
Orchestrated by GitHub Actions, powered by CrewAI agents.
A GitHub Actions pipeline sequences the work step by step—code review, merge, deploy, production test, rollback, DevOps—and each step runs a CrewAI agent defined the same disciplined way. That split keeps the system easy to reason about, govern and audit: new capabilities are added by dropping in a new agent, not by rewriting the pipeline.
◈ Anatomy of a Req2Prod agent
Agent · role + goal
Who it is and what it's accountable for—e.g. "Software Engineer: implement the task in clean, tested Python."
Backstory · persona priming
Context and expertise that shape how the agent reasons and the standards it holds itself to.
Task · description + expected output
The concrete unit of work and its acceptance criteria, assigned and tracked by the orchestrator.
Memory · short + long term
Recall of the project context, prior decisions and past runs—so the crew learns and stays consistent.
Skill · tools it can call
The real actions—run code, execute tests, use git, query docs, deploy—that turn reasoning into results.
LLM — set per agent in a registry swap the brain, keep the role
Haiku for cheap mechanical steps, Opus where the stakes are high, Sonnet by default—tuned role-by-role.
⚙ One model per agent — set in a registry
Each agent's model is set independently in a model registry (config/agent_models.json), with separate API and self-hosted "subscription" tiers. Changing a model is a one-line change.
Agent
Model (API tier)
Product Managerscopes intent, flags ambiguity
claude-sonnet-5
Software Architectplans & designs the change
claude-sonnet-5
Software Engineerwrites the code, opens the PR
claude-sonnet-5
code_reviewergates every merge
claude-sonnet-5
pr_fix_agentfixes review findings
claude-sonnet-5
pr_arbiterfinal merge / hold decision
claude-opus-4-8
prod_testersmoke & health checks in prod
claude-sonnet-5
devops_agentdiagnoses & fixes deploy failures
claude-sonnet-5
rollback_agentrestores previous commit
claude-haiku-4-5
Real assignments from the registry: Haiku for mechanical steps, Opus for high-stakes decisions, Sonnet as the default. Eight further agents (security, UX, and a CTO-led tech-excellence panel) round out the crew.
▣ No lock-in
Model-agnostic architecture rides the frontier as it moves and negotiates on price—never trapped on one vendor.
✓ Enterprise-ready control
An independent merge arbiter, automatic rollback, and a circuit-breaker on the DevOps agent keep it safe for regulated buyers.
⤢ Extensible by construction
New agents slot into the same GitHub-Actions + CrewAI structure—fast to adapt to new stacks and domains.