A personal concept: agents run the pipeline, a human signs every PR.
Why the pipeline can move fast without dropping quality: agents do the repetitive work; a person makes every decision.
The agent pipeline
Five specialized roles that hand work off to each other, with checkpoints where a human approves or corrects.
These five roles are the visible surface. Behind them runs a wider pipeline — strategist, architect, SEO, CI/CD, QA, monitoring — 14 specialized agents in total.
- 01Turns the brief into an implementation plan with atomic tasks and acceptance criteria.
- 02
Designer
Generates wireframes and design system, then a clickable Claude Design prototype. Always passes through human review. - 03
Coder
Writes code in small PRs. Follows repo conventions and stack patterns. - 04
Tester
Generates unit, integration, and E2E tests. Runs the suite in CI before requesting review. - 05
Security reviewer
Audits each PR against OWASP top 10, dependencies, and leaks. Flags risks for humans.
Where I come in
My commitment on this concept: I sign off four checkpoints in every build — spec, architecture, first PR, and release. No PR lands on main without a human pair of eyes.
- 01
Spec approval before any code is written
- 02
Architecture review before the first commit
- 03
Human code review on the first PR of every critical module
- 04
Light pen-test and release sign-off
What the pipeline targets
- WeeksTarget velocity
delivery cycles measured in sprints, not quarters
- Fixed scopeDelivery model
each sprint has a defined scope and deliverable — no open-ended retainers
- 100%Human review per PR
a hard model commitment: a human signs every change
Frequently asked
- Is the quality comparable to a traditional agency?
- Yes. Agents generate more code, humans review with more time. Test coverage and lint strictness usually end up tighter than in human-only projects.
- If this were a real engagement, who'd own the code?
- The client. The methodology hands over the repo, pipelines, and docs — no lock-in to internal tooling.
- In a real engagement, would data go through third-party models?
- By default the pipeline would use hosted models (Anthropic, OpenAI) under no-training policies. For sensitive data, private deployments (Bedrock, Azure OpenAI, or open-source models in your cloud). For LGPD- or Habeas Data-regulated data, the full stack on-prem — Ollama, Postgres, and Phoenix inside your network, no traffic to hosted models.
- Who are the humans in the loop?
- On this concept, it's me — Marian Sanjur, the designer and engineer behind Solaar. I review every checkpoint in the pipeline. In a real engagement, additional senior engineers would follow the same review model.
See the pipeline in action
Explore a real build: the agent logs, and the points where a human stepped in.