The agent writes the code. The engineer owns the outcome
Pini Shvartsman — Technology Leader · AI Transformation & Autonomous EngineeringI redesign engineering organizations around AI-powered execution and human accountability — and I build the autonomous systems that make it real.
Scroll
Execution moved.
I lead engineering organizations through this shift. Here’s what actually changed.
“Writing code was never the bottleneck.”
The full argument →
“AI agents aren’t tools you use. They’re workers you manage.”
Agents on the org chart →
The development sequence: agents take every step; the engineer directs, judges, and owns from above
| Step | Executor before | Executor now |
|---|---|---|
| Plan | Human | Agent |
| Scaffold | Human | Agent |
| Write | Human | Agent |
| Validate | Human | Agent |
| Ship | Human | Agent |
| Operate | Human | Agent |
Cheap to produce. Not cheap to own.
The bill lands on whoever owns production. That seat is mine.
“The cost doesn’t show up at generation time. It shows up at ownership time.”
The economics →
“Engineering discipline became a career risk. That’s the shift.”
What actually broke →
“Throughput is legible. Judgment is invisible.”
The scarce asset.
I came up through every rung — support, NOC, full-stack, architecture. The engineers I lead won’t climb that ladder, so I’m designing the new one.
“The first group learned with AI as a teacher; the second learned with AI as a tool.”
Why I’m worried →
Marcus
247 commits last month. 23 features shipped. Velocity charts trending up. Then one question — “why did you structure the caching layer this way?” — and silence.
He accepts what looks good on the surface and ships fast. For six months, that looks identical to mastery on paper. Then the architecture decisions come home.
Sarah
Same company. Nearly identical velocity. Asked the same question, she walked through the trade-offs, the two AI suggestions she rejected, and the monitoring she added for the failure modes she predicted.
She interrogates the suggestions. Her code doesn’t cause incidents — and when it breaks, she diagnoses it in minutes. Her understanding is deepening, not thinning.
Next · 04 Overwatch ↓“The agent has all the answers. You are the only one who knows which questions to ask.”
The hopeful version →
Supervision is architecture.
Agentic Overwatch is the discipline I named — and the one I run.
“You cannot govern a workforce that runs flat out, around the clock, with a team that logs off at 5 PM.”
Agentic Overwatch →
The overwatch tier model
Watch everything, all the time. Surface what matters.
Reproduce, propose, fix — inside the guardrails.
Authorize what ships. Own what happens next.
Your presence lights the human seat Tap the third lane to take the seat
| Tier | Who runs it | What it does |
|---|---|---|
| T1 Detect | Agents | Watch everything, all the time; surface what matters |
| T2 Remediate | Agents | Reproduce, propose, fix inside the guardrails |
| T3 Judge | Humans | Authorize what ships; own what happens next |
“The diff is the claim. The evidence is the proof. The human is the judge, not the fact-checker.”
The evidence gate →
“They’re debating tools while we’re deploying workers.”
How my org runs →
Not a proposal — the operating model behind these notes, running in production.
Written from the seat.
Technology leader, hands still on the system. Everything above runs in production — I write down what holds.
Taking things apart since age eight. IRC channels, scripts, communities that didn’t care how old you were as long as you kept up.
No DevOps, no infrastructure, barely a product. Built all of it — the CI/CD, the platform, the offshore team — and grew from five people to leading engineering at global scale.
“I just kept saying yes to whatever scared me most.” The current answer: redesigning how an engineering org runs when agents do the executing.
“We were the failsafe, and the failsafe doesn’t get to sleep through the incident.”
A year inside a NOC →
How I work
Embed
Inside your team — your codebase, your incidents, your constraints.
Ship one
The first system goes to production, not to a slide.
Prove it
Measured against the work it replaced: real time, real budget.
Make it repeatable
The version that works becomes a capability your org owns.
The words I had to coin.
New work needed new names. Each one is argued for in a note — take them, use them, tell me where they’re wrong.
Supervising an agent workforce as a designed, tiered, shift-based discipline — not an afterthought.
The Agent Operations CenterThe control room for your fleet: agents run detection and remediation, humans hold judgment.
The Overwatch Maturity ModelFive levels from blind trust (L0) to an orchestrated, supervised fleet (L4).
The evidence gateMachines produce the proof — repro, tests, validation. The human judges the case, not the diff.
Cheap to produce, not cheap to ownGeneration is nearly free. Ownership — comprehension, maintenance, incidents — is where the cost lands.
Agents on the org chartAgents doing real work are workers: a named owner, boundaries, onboarding, KPIs.
Operators vs. AI-augmented engineersOne group learned with AI as a teacher, the other as a tool. Only one compounds.
Scaffolding vs. substitutionEvery prompt is one or the other: building your understanding, or replacing it.
The three bucketsEvery task sorted before automation: AI alone, AI with human approval, or forbidden.
Securing IntelligenceMy series on defending AI systems: prompt injection, supply chain, and the culture underneath.
Field notes
Everything I’ve published from the seat — 70 field notes on running engineering when agents do the executing and I stay accountable for the outcome.
Latest
- A Ticket Is a Lossy Compression of Intent
- Rethinking the SDLC: Execution Is No Longer the Constraint
- Everyone Trained Their Engineers on AI. The Gap Didn't Move.
- Make the Cheap Path the Default. Make the Expensive Path Prove It.
- Alberta Scanned 466 Million Lines of Code in 20 Hours. The Architecture Is the Story.
- Stop Reviewing Code. Start Reviewing Evidence.
- Agentic Overwatch: Why Your Next Dev Team Will Look Like a NASA Control Room
- The One-Man Show Company. Don't Let the Monkeys Touch Production.
- 100 Days to the EU AI Act Deadline. Your Engineering Team Hasn't Started.
- The Vibe Coding Backlash Is Right. Seniors Are Losing the Argument Anyway.
- The End of Courses: Learn From AI Like a Toddler, Or Become Obsolete
- I Don't Put All My Eggs in One Basket. Anthropic Is Making That Hard.
Not advice from a deck. Notes from the seat — while the system is still running.
I’m in the middle of this every day: building AI systems that investigate bugs, review code, and run operational workflows. If you’re somewhere in the same transition, I’d rather compare notes than watch you rediscover it the expensive way.
Let’s talk →