The Team.
Nineteen specialized AI agents across two workspaces — one that decides what is worth building, one that takes it all the way to a published, live site. Structured handoffs and quality gates between every stage.
Two workspaces, one organization.
Delivery and intelligence were one repo until July 2026, when they split — deliberately. They answer different questions on different clocks: delivery asks 'is this built and live?', intelligence asks 'is this worth building?'. Splitting them stopped a research cycle and a build pipeline from contending for the same context. They stay joined by one artifact — initiatives.md — where a research bet gets promoted into delivery via /create-prd or /create-frd.
How the organization is arranged.
Three tracks across the two workspaces. Delivery runs here and owns everything from intake to go-live. Research & CI and Product Management run in the sibling intelligence workspace — same agent-identity model, same orchestration layer, their own repo and their own clock.
One engine, composed rather than forked.
Domain knowledge — what you build on, where the spec comes from, who is driving — used to be spelled inline across the shared commands. That works until there is a second of anything, and then adding one means editing every stage. So a run now answers three independent questions at intake and the answers COMPOSE: three dimensions, ten runnable combinations, one engine. The incoherent combinations are refused by name at resolution time rather than discovered halfway through a build. Adding a platform is writing one declaration file; it is not a new flow, and it is never a new repo.
From idea to production.
Seven agents in sequence: Team Lead orchestrates, Product Manager specs, Architect designs, UI Designer proposes, Lead Developer breaks down, QA enriches, Developer implements. A final /ship stage closes the loop with a release checklist. Each output becomes the next input through structured handoffs.
Called when needed.
Six agents that operate outside the linear core pipeline. Four are pulled in when the context calls for them — to surface hidden assumptions before planning, ground a fuzzy claim mid-conversation, extract structure from an existing site, or run a heavy documentation pass. Two more are contributed by a platform pack rather than by the core: a DevOps Engineer that provisions infrastructure, and a Content Author that writes and lands content. They live inside the pack, and a run that does not compose that pack never sees them.
The builders.
13 agents that take a raw idea through brainstorming, specification, architecture, design, planning, testing, implementation, and shipping. 7 form the linear core pipeline; 6 are on-demand specialists — Rubber Ducky, Research Sidecar, Doc Writer and Source Analyst from the core, plus a DevOps Engineer and a Content Author contributed by a platform pack — invoked when the context calls for them.
Five stops, and you can remove all five.
Autonomy here is a dial, not a posture. There are exactly five places the pipeline stops for a human — and each one is a token you can untick independently. Tick all five and you get a supervised run. Untick all five and it is fully autonomous: even the live publish goes unattended. The point of naming them is that the set is finite and inspectable; you always know exactly where you will be asked.
Rigor is a question budget.
The spec stages are interrogations, not questionnaires. One question per turn, each derived from the last answer, pushing back on every hedge — run inside the spawned Product Manager against an on-disk ledger the next agent reads by path. The rigor dial sets the budget: light 3–5, standard 8–12, thorough 20+ and uncapped. It is a floor, not a target. A grill ends when the budget is met, when the questions stop finding anything, or when the operator takes the standing 'Enough — stop asking' option.
The researchers.
3 specialists producing vendor-neutral intelligence briefs. The Research Specialist covers technology topics across the modern web stack, the CI Specialist profiles competing platforms with factual SWOT analysis, and the DevX Specialist distils a developer-experience industry trend brief. All apply four shared lenses: release tracking, security monitoring, ecosystem trends, and community voice — and the track outputs strategic initiatives, not a task backlog. Lives in the sibling intelligence workspace since July 2026.
The product managers.
3 specialized PMs that evaluate managed repositories as product surfaces. Each consumes upstream research and CI briefs, inspects the actual repo, and produces a register of initiatives — JPD-altitude bets gated by altitude and control — rather than a granular task backlog. A chosen initiative is promoted into delivery via /create-prd or /create-frd.
Go deeper into how this gets shipped
Methodologies and case studies show the agentic pipeline in motion — the patterns the agents converge on, and what that produced when run end-to-end against real systems rather than mocks.
Build your own agentic workflow
This framework is extensible by composition rather than by forking. The same agents adapt to any domain through command injection and a declarative pack — never a prompt rewrite, and never a second copy of the engine.