AI-Native Product Studio · Playbook
Playbook
How we operate: the pipeline, the agents, the guardrails, the metrics. This is the public documentation of our studio model — transparent by design.
Operating Philosophy
Delegate, Review, Assume
Three words that define every interaction between humans and agents in our studio. The model is simple; the discipline it requires is not.
Delegate
Every well-defined task goes to a specialized agent. If a task can be expressed as a clear prompt with acceptance criteria, a human should not be the one executing it.
- → Code generation, test writing, documentation
- → Research, competitive analysis, risk assessment
- → Refactoring, security scanning, cost monitoring
Review
Human validates every output against technical and business criteria. No agent output reaches production without explicit human approval.
- → Does it meet acceptance criteria?
- → Does it align with architecture decisions?
- → Does it introduce risk the agent may not see?
Assume
The human signs the merge. Responsibility is always human. Agents are tools with extraordinary capabilities, but accountability cannot be delegated to a model.
- → Human merges to main, always
- → Human owns the decision, not the agent
- → Human communicates with the client, always
The Pipeline
From idea to production in four phases
Every project follows the same pipeline. The phases overlap, the agents change, but the structure is constant.
Discovery
Days 1–3
Human Role
Client Partner conducts discovery conversations, understands the problem space, captures constraints and success criteria. This is the most human-intensive phase — nuance, empathy, and business context cannot be delegated.
Agents at Work
Research Agent scans the competitive landscape, Domain Agent maps industry-specific terminology and data models, Risk Agent flags technical and regulatory risks early.
Specification
Days 3–7
Human Roles
Product Architect defines the product shape and user flows. Systems Engineer defines the technical architecture, data model, and integration points. Together they translate business intent into machine-readable specifications.
Agents at Work
Spec Generator produces user stories with acceptance criteria from conversations. Architecture Agent proposes system diagrams and component boundaries. Estimator Agent breaks stories into tasks with complexity scores and token-cost projections.
Build
Weeks 2–N
Human Roles
Systems Engineer orchestrates the agent fleet, reviews PRs, and makes architecture calls. Quality Curator validates test coverage, UX consistency, and acceptance criteria. This is where the agent leverage is highest — 7+ agents working in parallel, human reviewing outputs.
Agents at Work
Agents work in parallel branches. Backend generates API endpoints and business logic. Frontend builds components and views. Database writes migrations and seeds. Test Agent generates unit, feature, and integration tests. Doc Agent keeps documentation synchronized. Security Agent runs Trivy, OWASP checks on every push. Refactor Agent identifies and resolves code smells.
Deploy & Operate
Continuous
Human Role
Reliability Engineer owns uptime, approves deployments, manages incident escalation, and reviews cost trends. The goal is zero-downtime delivery with full observability.
Agents at Work
Deploy Agent manages CI/CD pipelines and rollout strategies. Observer Agent monitors metrics, logs, and traces in real time. Incident Agent triages alerts, proposes root cause, and drafts incident reports. Cost Agent tracks token spend, cloud costs, and flags anomalies before they become problems.
The Numbers
Traditional model vs. Studio model
Side-by-side comparison for a typical mid-complexity SaaS product. Numbers are based on real projects.
| Metric | Traditional | Studio Model |
|---|---|---|
| Team size (humans) | 7 | 4 |
| Agents in fleet | 0 | 10 – 25 |
| Velocity | 30 – 40 sp/sprint | 60 – 100 sp/sprint |
| Lead time (story to deploy) | 3 – 5 days | Hours |
| Monthly cost | Alto (equipo grande) | Menor (equipo reducido + agentes) |
Story points are illustrative. Actual velocity depends on codebase complexity, test coverage requirements, and domain specificity. Token costs assume Anthropic Claude as the primary model; costs vary with model mix and prompt engineering efficiency.
Guardrails
What we never delegate
Velocity without guardrails is just chaos. These are the hard rules that no agent, no matter how capable, is allowed to bypass.
Architecture decisions
Any change affecting more than 3 files requires explicit human review before merge. Structural changes are never auto-merged.
Production DB schemas
Database schema modifications in production are always human-reviewed and human-executed. Agents can propose migrations; they cannot run them.
Credentials & secrets
No agent has access to production credentials, API keys, or secrets. Agents work with environment stubs and mock values only.
Client communications
All client-facing communications are written and sent by humans. Agents can draft, but a human always reviews and sends.
Legal, commercial & pricing
Legal decisions, contract terms, pricing structures, and commercial agreements are exclusively human domain. No exceptions.
Merge to main
No agent can merge to the main branch. Every merge requires a human reviewer who assumes full responsibility for the change.
Enforcement — These guardrails are not guidelines. They are enforced through branch protection rules, CI pipeline gates, MCP tool permissions, and mandatory human approval steps. Violating a guardrail stops the pipeline.
Our Stack
home.stack.title
We are opinionated about our tools. We pick few, learn them deeply, and squeeze every ounce of leverage from them.
home.stack.we_master
home.stack.we_use
home.stack.we_explore
Raccontaci cosa stai costruendo.
You have read the playbook. You know how we work. If this resonates with how you want your next product built, let's talk.