Zack Babtkis

I’ll get your team shipping production software with AI agents. I’ve actually done it.

At Pie, a company with two million users, I built and shipped a production product where AI agents wrote the entire codebase. It wasn’t a prototype. It made money, I owned it end to end, and the agents worked inside a harness I engineered. Most AI consultants learned this from a course. I learned it by shipping.

Book a scoping callCurrently booking September 2026
100%of the codebase agent-built on the product I shipped at Pie
2M+users at the company where it shipped as a growth engine
Dailyagent-driven development is how I work, not a demo I give
12+ yrsengineering standards to hold the output to (ex-PayPal Senior Staff)

The proof

Case study: pie.yt

In 2025 I built pie.yt, an ad-free YouTube viewer, as Pie’s first product developed entirely through what we called harness engineering: instead of writing the code, I engineered the environment — the conventions, context files, guardrails, and review gates — that let AI agents write production code I would sign my name to. I owned it from conception to launch.

It shipped, it worked, and it became a growth engine for a two-million-user company. The code met the bar I spent five years holding as a Senior Staff engineer at PayPal, because the harness required it to. That harness is what I build for your team.

Sound familiar?

  • Everyone has a Claude Code or Copilot license; your cycle time hasn’t moved.
  • One or two enthusiasts get real results and nobody can replicate what they’re doing.
  • Agent-written PRs keep failing review, so seniors quietly went back to writing everything by hand.
  • Nobody has decided what agents may touch, so security blocks everything by default — or worse, nothing.
  • Costs are climbing with no way to tell productive token spend from waste.
  • Leadership is asking for an "AI strategy" and what you have is a pile of subscriptions.

There's a long distance between installing the tool and running it well across a team. Crossing that distance is the engagement.

What agent-ready actually looks like

Agents usually fail because the codebase and workflow give them nothing to hold onto, not because the models are weak. Here is what changes:

Before: agent-hostile

  • Conventions live in senior engineers’ heads
  • Thin or slow tests, so nothing verifies agent output
  • No CLAUDE.md, no context files — every session starts from zero
  • Agents get blanket access or no access; security is a standoff
  • Each engineer prompts differently; results can’t be reproduced
  • Token spend untracked and unbounded

After: agent-ready

  • CLAUDE.md conventions and skills encode how your team builds
  • Fast test gates that agent work must pass before review
  • Hooks enforce policy automatically — no relying on the agent’s memory
  • MCP wiring gives agents your internal tools, scoped and audited
  • Subagent patterns for review, testing, and migration work
  • Cost controls and metrics: spend per merged PR, not per seat

What your team gets

Every item below gets committed to your repos and documented.

  • Harness configuration for your actual repos — CLAUDE.md conventions, skills, hooks, and guardrails
  • MCP integration — your internal tools, docs, and services wired in so agents work with your context, not generic knowledge
  • Security and permissions policy — what agents may touch, enforced technically (scoped credentials, sandboxing, allow-lists, human gates on irreversible actions), written with your security team
  • CI integration — test and review gates that agent-written work has to pass before a human reviews it
  • Cost controls — budgets, tracking, and the metrics that distinguish productive spend from waste
  • Real merged work — we ship items from your actual backlog together; the first harness-engineered feature is the training vehicle
  • Playbooks your team keeps — onboarding docs and patterns so the practice outlives the engagement

Start small: find out what's actually in the way

Agent-Readiness Audit

$3,000

One to two weeks · fixed price

  • Codebase evaluation through an agent’s eyes: tests, docs, structure, CI — what blocks reliable agent work today
  • A live pilot: I run a real task from your backlog through an agent harness on your actual code, and you watch
  • Starter harness config for your repos (CLAUDE.md, conventions, guardrails, review gates)
  • Security and permissions recommendation your security team can sign off on
  • A sequenced rollout plan — who starts, on what work, measured how — with a fixed quote for the embedded engagement

If we continue to a full engagement, the entire fee is credited toward the project.

Start with the agent-readiness audit

Engagements — priced per engagement, never hourly

  1. Agent-Readiness Audit · $3,000 fixed

    The entry point above. Useful even if you run the rollout yourselves — and the fee is credited if we continue.

  2. Embedded Enablement · 2–4 weeks, fixed quote

    I set up the harness across your repos and ship real backlog items with your engineers pairing alongside. Your skeptics will come around when they review the merged PRs. Quoted from the audit.

  3. Fractional AI Engineering Lead · monthly retainer

    The tools change monthly; someone senior has to own the practice. A flat retainer keeps the harness current, reviews agent workflows, and levels up each team as it onboards.

When this won’t work

An honest list, because this category is drowning in hype. I will tell you in the scoping call — and again in the audit — if you’re in one of these situations, and I’ll decline the engagement rather than take money for a rollout that won’t stick:

  • Leadership wants a headline, not a workflow change. If nobody senior will spend time in the new workflow, the tools will sit unused no matter how well they’re set up.
  • The codebase can’t verify anything.If there are close to no tests and no appetite to build a minimum verification layer, agent output can’t be trusted at scale. Fixing that comes first, and it may be all you need from me.
  • You’re hoping to replace engineers wholesale. This multiplies good engineers. It doesn’t replace them.
  • Compliance forbids code leaving your network and you can’t run approved models. Sometimes there’s a path. Often the timing is just wrong. I’ll tell you which it is.

Questions clients ask

Which tools do you set up?

I’m deepest in Claude Code and the Claude Agent SDK — the stack I used to ship pie.yt and the agent infrastructure I built at ZeroClick. The harness itself (conventions, context, hooks, gates, CI) is deliberately tool-agnostic, which is what protects the investment when the tool landscape shifts next quarter.

Will agent-written code pass our review bar?

That’s the defining constraint of the engagement. The harness includes the review gates, testing requirements, and conventions that make agent output meet a senior engineer’s bar — mine was a PayPal Senior Staff bar. Where the codebase blocks that, fixing it is sequenced into the plan.

Is this training or consulting?

Neither, mostly — it’s embedded engineering. We ship your actual backlog together and the workflow transfers by doing. The first harness-engineered feature your team ships is the training vehicle.

How do we measure whether it worked?

We pick metrics up front — typically cycle time on the pilot team, PR throughput and revert rate, cost per merged PR, and the share of merged work that’s agent-authored. You should see movement during the engagement, not after it.

Our security team is nervous. What do agents get access to?

Whatever you decide, enforced technically: scoped credentials, sandboxed execution, allow-listed commands, and human gates on anything irreversible. I write this policy with your security team, not around them.

Other things I’m hired for

Chrome → Safari & iOS Extension Porting

I port Chrome extensions to Safari, macOS, and iOS — the way I did at Honey and Pie.

Learn more →

MCP Server Development for SaaS

I build production MCP servers that put SaaS products inside Claude and ChatGPT.

Learn more →

Find out what's actually in your team's way.

Book a scoping call. Tell me your team size, your stack, and what you've tried. I'll give you a straight answer, including 'you're not ready yet, and here's what to fix first.'

Email me about your projectI personally reply within one business day. · Currently booking September 2026.