中文
AI Engineer World's Fair

Recognize the worse version of CI/CD you're rebuilding by hand, and stop it with five gates

Every Solo Agent Builder Eventually Reinvents a Worse Version of CI/CD - Sumaiya Shrabony · Sumaiya Shrabony

11 min
AgentEvals

11 min total·Actually worth watching closely: ~4 min·3 must-watch clips

Orange = the 4 minutes worth watchingFor the rest, the guide is enough
Segment guide · 8 segments
  1. 0:00 0:51Listen

    What you're rebuilding has a name

    The opening claim, stated flat out: build agent systems alone for long enough and you start building something that looks suspiciously like CI/CD — except worse, because you build it from scratch, one failure at a time.

    The problem isn't that you lack tools; it's that you're already building them without realizing it, which is why you build them badly.

    Pure spoken argument over a title slide — fine to listen to while doing something else.▶ Jump to 0:00
    Speaker · Sumaiya Shrabony
  2. 0:51 2:30Watch

    First, the system that taught her

    The speaker shows the agent pipeline she actually runs: which agents sit between input and final artifact, and where each hands off to the next.

    Every single handoff is a place where the system can lie to you.

    The full pipeline structure is on screen — where the arrows point and where they break only works if you're looking, and all three failures later come back to this map.▶ Jump to 0:51
    Speaker · Sumaiya Shrabony
  3. 2:30 4:24Listen

    The rebuild order is predictable

    Walks the fixed path solo builders take: a test for whether output still matches the expected shape, then alerts on the scheduled task, then validation at the boundary, a checkpoint before the ready folder, and logging everything.

    The order isn't a coincidence — agent systems give you none of these operational guarantees by default, so the pain arrives in the same sequence for everyone.

    A paced, list-like stretch of argument; the audio carries all of it, no need to watch.▶ Jump to 2:30
    Speaker · Sumaiya Shrabony
  4. 4:24 5:18Watch

    The dangerous failure isn't a bad output

    Points out that agent demos mislead because they always show the happy path, then shows an artifact that looks entirely ready and shouldn't ship.

    Your agent will fail. The real problem is a system that formats that failure nicely and ships it downstream.

    The key moment of the talk: every section of the artifact is there and the status reads ready — the flaw only shows on screen, and no description conveys that "nothing looks wrong" feeling.▶ Jump to 4:24
    Speaker · Sumaiya Shrabony
  5. 5:18 6:28Skim

    Failure one: voice drift

    The first real case — generic AI marketing copy that gets saved because the sections are complete and the status is fine. The fix isn't to improve it, it's to block it at the boundary.

    For a content system, the voice contract should be your first gate, and its job is to refuse, not to repair.

    The example text sits on a static comparison slide; one glance at the tone of that boilerplate is enough — no need to read the whole screen.▶ Jump to 5:18
    Speaker · Sumaiya Shrabony
  6. 6:28 7:29Watch

    Failure two: a claim with no source

    The second case — a specific, plausible number like "reduces rework by 37%" with nothing behind it: check the verification log and it's empty.

    Once your system makes claims about data or users with no validation chain, what you publish is an unverified assertion in a professional-looking wrapper — a credibility problem, not an agent problem.

    The precise number and the blank verification log are placed side by side; the force of that contrast is entirely visual, and listening alone cuts it in half.▶ Jump to 6:28
    Speaker · Sumaiya Shrabony
  7. 7:29 8:41Skim

    Failure three: the recurring opening angle

    The third case, and the most realistic one for solo builders — the output is new and technically coherent, but the opening angle is a near duplicate of past output.

    Your audience notices the repetition before you do, so you need a duplication check against your vault history.

    Mostly a list of past hooks on screen; skim until the repetition registers — nothing here needs a pause.▶ Jump to 7:29
    Speaker · Sumaiya Shrabony
  8. 8:41 10:51Listen

    Five boring gates, and the first step

    Closes with something you can act on: an output contract, a voice contract, a verification contract, a duplication check, an audit trail — plus the first step, mapping every handoff and putting the gate at the boundary where bad data costs the most.

    A gate that only logs warnings isn't a gate, it's a suggestion — it has to say no and actually block the artifact. Before you add another agent, add one boundary.

    The ending is pure method, delivered standing; writing these five down matters more than watching the screen.▶ Jump to 8:41
    Speaker · Sumaiya Shrabony