Phoenix is a Elixir natural fit for agent product. GreatCTO auto-detects both — adds the agent-product archetype overlay, wires agent-product-specific gates, and runs 83 specialist agents around your existing Phoenix workflow.
GreatCTO reads your mix.exs and detects phoenix + agent-product archetype from signals: imports, file structure, env vars, README hints.
Attaches the agent-product archetype overlay: OWASP LLM Top-10 evals, prompt-injection review, EU AI Act high-risk classification. Override if your specifics differ; the defaults are sensible for Phoenix-style projects.
qa-engineer runs mix test --cover / credo --strict; security-officer audits Ecto query injection + Phoenix CSRF; performance-engineer profiles BEAM scheduler + Ecto pool.
Bugs you've hit before in other Phoenix projects (connection-pool exhaustion, ORM N+1 queries, retry storms) — the agent's Step 0 includes the prior detection order. MTTR drops 94 % on second occurrence (methodology).
$ cd my-phoenix-app && npx great-cto init ✓ scanning manifests… found manifest ✓ stack: phoenix (Elixir) ✓ archetype: agent-product ✓ overlay: applied ✓ 83 agents ready $ /start "add tool-calling endpoint for the agent loop" ▸ architect drafting ARCH-agent-product.md… ▸ pm decomposing into beads tasks… ⚐ gate:plan — your approval needed
Approve → 3 senior-devs run in parallel worktrees → 5 reviewers fan out in parallel → gate:ship → deploy. One real run walked stage-by-stage: /proof.
This is the shape of what senior-dev drafts for "tool-calling endpoint for the agent loop" — auth first, schema validation, and the audit line the agent-product reviewer requires before gate:ship opens.
# lib/app_web/controllers/agent_product_controller.ex — reviewed by 5 agents
defmodule AppWeb.AgentProductController do
use AppWeb, :controller
plug :require_authenticated_user # security-officer: auth before handler
def create(conn, params) do
with {:ok, attrs} <- validate(params), # qa-engineer: changeset enforced
{:ok, result} <- AgentProduct.handle(attrs, conn.assigns.current_user) do
AuditLog.record(conn.assigns.current_user.id, "tool-calling endpoint for the agent loop", result.confidence) # gate:agent-product
json(conn, result)
end
end
endagent-product overlay.Customer-support agents with tool-calling and human escalation.
Research / browsing agents with cost-runaway protection.
Workflow copilots embedded in existing SaaS products.
No black-box "AI does it all" loop. GreatCTO is a deterministic state machine — 8 stages, 22 nodes, 2 human gates. Every node maps to a real agent on GitHub. Inspect the state machine →
$ npx great-cto init
Free, MIT, runs locally. Built as a Claude Code plugin — install with one command.
Eight stages, two human gates, four memory layers. Why this exact shape, and what I tried that didn't work.
One run, one feature, from prompt to merged PR. Time, cost, and gate-by-gate breakdown — no marketing math.
Regex vs LLM-based archetype detection, the false-positive count, and why I keep rejecting the obvious fix.
The bottleneck in agentic SDLC isn't model quality — it's process governance. Here's the state machine that closes the gap.