// FDLC · How a change ships

One card becomes a verified outcome.

Our agents take a card from the board to production through nine lifecycle stages. SOSA™ sits around every stage: each one is supervised, scoped and logged, and money, client messages and new outreach always wait for a human yes.

Watch an example run See the nine stages
Built on its own branchReviewed on the exact commitTested on stagingTested again on prod
// The lifecycle

Nine stages. SOSA around every one.

The stage names come from FDLC, the Factory Development Lifecycle. What OpsAgents AI adds is who does each stage, and the SOSA control that holds while it runs.

SOSA™ around every stage · Supervised · Orchestrated · Secured

Inside each stage, the agent runs SOSA’s loop: Plan, Act, Verify. Every action is logged to the card, and each stage hands over only when its check passes.

  1. Intent

    PM agent

    A card on the board, with acceptance criteria written by the PM agent.

    SOSA controlAn unclear card is refused, with the reason written on the card.

  2. Planning

    Routine + engineering agent

    The board pack knows the product. The routine claims the card and the engineering agent picks the files the change touches.

    SOSA controlScoped access through the gateway. No secrets in code.

  3. Execution

    Engineering agent

    It writes the change and its tests, then opens one pull request for one card.

    SOSA controlIts own branch, never main. It never merges its own work.

  4. Evidence

    CI + QA agents

    CI runs on that exact commit, and QA agents test it on staging in a real browser.

    SOSA controlGuards are probed both ways, allow and refuse. Test data only.

  5. Verification

    CTO agent

    The CTO agent reviews the exact commit that will ship, and can send it back. Human code review is your choice.

    SOSA controlNo merge without the review, pinned to that commit.

  6. Approval

    A human

    Money, client messages and new outreach wait for a human yes. Everything else flows.

    SOSA controlThese three always stop for a person.

  7. Release

    Gateway

    The gateway deploys the tested commit to Google Cloud, and reads back what is live.

    SOSA controlDestructive commands are blocked at the gateway.

  8. Outcomes

    QA agents

    A prod check and the regression suite run right after release. The card is Done only when they pass.

    SOSA controlAn audit trail on the card. A failure rolls back and reopens it.

  9. Learning

    Every agent

    Every incident becomes a rule in the handout that every agent reads first.

    SOSA controlEach incident becomes a rule.

Lifecycle stages from FDLC, an open framework by fdlc.ai. FDLC is not ours. The agents, gates and SOSA controls shown are how OpsAgents AI runs each stage. Read how SOSA works.

// One run, start to finish

Agents build, test and ship. SOSA on every step.

One ticket, from the board to live: built, reviewed, tested on staging, shipped, and tested again on prod. Step through it, or let it play.

EXAMPLE RUN · illustrative data

    AutoPilot · board → PRTo Do
    Add a “talk to a person” hand-off

      Acceptance criteria PM

      1. A customer can ask for a person
      2. The support team gets it on WhatsApp
      3. Conversation context goes with it
      4. No hand-off outside business hours

      Files it will touch ENG

      • app/routes/handoff.ts
      • app/lib/whatsapp.ts
      • app/lib/hours.ts
      • tests/handoff.test.ts
      Scope check: clear enough to build. Unclear cards are refused, with the reason on the card.Planned
      Pull request openedone PR for one card · reasoning attached · its own branch, never main
      Awaiting review
      • CTO
        Code review of the exact commitautomatic; it can send the work back
        ✓
      • CI
        Full test suite and evalsbehaviour evals and the test harness, on this commit
        ✓
      • YOU
        Human code review: your choiceon at the start; automatic once evals and the harness earn it
        ○
      • SOSA
        Gate checkmoney, client messages, new outreach: always a human yes
        ✓
      Staging is serving this exact commita deploy is checked by reading what is live, not by a green build
      On staging
      • QA
        Every acceptance criterion, in a real browserdesktop and mobile, front end and back end
        ✓
      • QA
        The control: the old build fails the same testso the test proves the fix, not just a green tick
        ✓
      • SOSA
        Test data onlyno real customer is messaged by a test
        ✓
      Real runs, not this example
      • 15,114tests passed across 878 test files in about 3½ minutes, on one CI run of our Maya product (11 Oct 2026)
      • 8/8QA checks passed on staging, the old build failed the same check, then 8/8 on prod (6 Oct 2026)
      • 94%of 4,355 merged pull requests were reviewed by the CTO agent, over the 30 days to 11 Oct 2026

      Source: our own CI runs, QA logs and repository history, read 11 Oct 2026.

      The card moves to Done only after prod passesa failed prod check rolls back and reopens the card
      Prod verified
      your-product.app/support
      Hi! How can I help with your order?
      The delivery date looks wrong. Can I talk to someone?
      Of course. I’m handing you to a person now, with this conversation attached.
      Talk to a person
      Hand-off to a person is liveshipped by the gateway · prod check passed right after
      Live

      Example data. This is an illustrative run of AutoPilot’s real flow (AutoPilot). The ticket, product screen, files, commit ids and log lines are example data. Only the “Real runs” figures in the QA step are measured, and each carries its date. Stage names from FDLC, an open framework by fdlc.ai, not ours.

      // Built · reviewed · tested on staging and prod · SOSA™ on every step

      Everyone is a vibe coder now: one prompt makes a demo, and that is fun. Delivering a robust, scalable, secured product is still hard, even with agents. It takes professionals.

      We architect so you can vibe.

      Scope your first build See AutoPilot