Code is cheap now. Builders aren't.

On the product I run alone, 97% of commits since July 2026 were written with Claude Code. Agents can't find out what people need or tell good from good enough. That's the work that pays from here on, whether you've shipped for two years or twenty. My whole setup is public, numbers included.

Get the free kit Read the free guides

Same job title, whole new job. See the numbers.

The software job, in the order agents take it over

Writing the codeAgents, already
Reviewing itAgents, partly
Shipping and scaling itAgents, next

The bottleneck stops here

Finding out what people needYou
Defining what good meansYou
Making it something people loveYou

Based on Laurie Voss's talk at AI Engineer.

My agents work nights

A scheduled agent runs OpenWhen every day. It reads the numbers, opens PRs, has a second agent grade them, and merges what a revert can undo. Anything that touches money waits for me. Grading started on 31 October 2026; before that a list of file paths decided what merged on its own. Here's the record so far.

PRs merged on OpenWhen since July 2026: 157
opened by the scheduled loop, the rest in sessions I directed: 114
revert on main since July 2026: 1
of 538 commits on main written with Claude Code: 97%

The charter, grader, hooks and scripts are free on GitHub. How each number is counted, and the one revert, are on the log.

One agent, one night (simplified)

21:00ops-agent starts, reads CHARTER.md
21:02sweep: 3 open PRs, rebased on main
21:14hook blocked: git push --force origin main
21:15pushed to branch fix/stale-cache instead
22:40PR opened: fix stale cache on pricing page
22:43reviewer subagent: grade A, tests green
22:44merged, deploy green
01:10PR touches checkout: grade X, left for a human
07:00morning report sent: 2 merged, 1 needs you

The job that's left

Writing code used to be what you got paid for. Agents do it now. What's left is the work around it, and it's the same at every level, whether you set direction as a staff engineer or you're still earning your first big project.

Find out what people need
No model knows what your users struggle with. The person with the problem does, and you get it by asking.
Write down what good means
Specs, rubrics, evals. Turn the taste in your head into a test an agent passes or fails.
Review only what can hurt
Let agents check what's cheap to check. Spend your own attention on the code that can take the business down.
Measure what ships
Watch what reaches people and what they do with it.

The course

Agent Ops teaches you to run this setup on your own repo. It's the part of the job you control even if you don't decide what gets built at work. Take it on your own, in a live group, or with your team.

Run an agent on your repo every night, with guardrails.

Agent Ops

For engineers who already ship and want Claude Code doing more of the work, with guardrails that catch its mistakes.

You finish with an agent shipping guarded, graded PRs in your repo on a schedule.

Write the charter your agent obeysWeek 1
Guardrails with hooksWeek 2
Run headless on a scheduleWeek 3
Skills and subagents worth keepingWeek 4
Evals and independent PR gradingWeek 5
Escalation and the kill switchWeek 6
Roll it out to a whole teamWeek 7

Your instructor

Albert Nahas

Senior engineer, ex-Miro. I train engineers to build products with AI.

I'm Albert. I've shipped software for 13 years, most recently at Miro and now at a European health-tech scale-up. I've put more than 5,000 hours into building with AI agents and trained more than 100 engineers to do the same.

My job changed under me. I barely type code now, and I ship more than before. On the side I built OpenWhen alone, and more than 70,000 people have written or opened a letter on it. Agents did most of the typing. What made it work was talking to the people using it and deciding what good looked like.

Most engineers I talk to, from new hires to principals, know the job is changing. I publish what mine looks like, numbers included. Everything I teach, I do every week, and the setup itself is open source as agent-ops-kit.

people have used OpenWhen, a product I built alone: 70K+
hours building with AI agents: 5,000+
engineers I've trained to move from coding to product building: 100+
years shipping software: 13

Pricing

Take Agent Ops on your own with a review of your setup, or join a live cohort and get feedback every week. Teams pay a flat fee per day, not per head. Nothing is for sale yet: every option opens from the waitlist. Every price comes with an invoice your employer can pay from a learning budget.

Self-paced

Agent Ops for one engineer, at your own pace

€249

one-time, excl. VAT

An exercise repo with a solution branch for every module
A short screen recording for each module
A written review of your own setup: charter, hooks, grader
A printable field guide
A completion record for EU AI Act literacy (Article 4)
Counts toward a cohort seat if you upgrade within 60 days
Team licence for up to 12 engineers: €1,490

Join the waitlist

Cohort seat

Agent Ops for one engineer, live with a small group

€990

founding cohort, then €1,490, excl. VAT

6 to 7 weekly live sessions, 2 hours, evenings CET
A group of 4 to 10 engineers, so everyone gets airtime
Weekly office hours and a private group
Written review of the agent you build in your repo
Feedback on your setup every week, not just at the end
Every session recorded
Full refund until the end of week 2
Invoice to your employer for your learning budget

Get the cohort dates

Team workshop

For one team, up to 12 engineers

€3,500

per day, remote or hybrid, excl. VAT

Call beforehand to fit the day to your codebase
Hands-on in your own repo, on real tickets
A team playbook: CLAUDE.md, hooks, permissions, PR policy
30 days of follow-up questions by email
Attendance certificates for EU AI Act literacy (Article 4)
Pilot: the first team pays half for a published case study

Ask about a date

Organizations

For teams above 12 engineers, or a rollout across your org

Let's talk

Priced to your setup

Everything in the team workshop
Three follow-up sessions, one a week
Guardrails and independent PR grading set up in your CI
Before and after numbers: cycle time, review load, reverts
60 days of follow-up

Talk to me

Questions

Who is Nightshift Dev for?

Working software engineers at any level who already code with an AI assistant and want agents doing more of the work, safely. Agent Ops runs on your own repo, so it fits whether or not you decide what gets built at work.

The kit is free. What does the course add?

The kit at https://github.com/albertnahas/agent-ops-kit is the setup itself, and the guides explain each piece. The course adds exercises with a solution branch each, the reasoning behind every rule, and my review of your setup in your own repo. You can build it from the free material alone, and plenty of people will.

I've never used Claude Code. Is this for me?

Not yet. Agent Ops assumes you already write code with an AI assistant. I'm weighing a Claude Code basics course for engineers who haven't used it yet, and it opens only if enough people want it. Join its waitlist below to count yourself in.

How much does it cost?

Agent Ops self-paced, with a review of your setup, is €249. A live cohort seat is €990 for the founding cohort, then €1,490. A team workshop is €3,500 a day for up to 12 engineers. Bigger teams and whole organizations get a price that fits their setup. Prices exclude VAT.

Does this work on a team, or only on a solo product?

So far it has run on products with one human: 157 PRs merged on OpenWhen since July, 1 reverted, all public at https://nightshiftdev.xyz/log. A team adds things a solo repo doesn't have, like review norms, compliance and many authors. That's why the first team workshop is a pilot at half price, in exchange for a published case study with before and after numbers.

What about The Future of Engineering?

Its guides are free now. The course, on building a product end to end, opens only if enough people ask for it. Join its interest list to count yourself in.

Can my employer pay?

Yes. You get an invoice addressed to your employer, so it can come out of your learning budget.

How much time does it take?

Six to seven weekly live sessions of two hours, in the evening, Central European Time. Office hours run every week too, if you want them. Groups are 4 to 10 engineers, and a cohort runs once 4 have signed up.

What if I miss a session?

Every session is recorded, so you can catch up the same week.

What does a team get?

A workshop day in your own repo, remote or hybrid, and a playbook your team keeps. The first team gets it at half price for a published case study. Organizations can add weekly follow-ups, guardrails set up in CI, and before and after numbers on cycle time, review load and reverts.

When does the first cohort start?

The first cohort is planned for mid-January 2027, once 4 people have signed up. Everything is on a waitlist until then. Join it and you'll get one email with the exact dates.

What if it isn't for me?

You get a full refund until the end of week 2.

Free guides

Free and detailed. Every guide is a setup I run myself, code included.

All guides

Agent Ops

The Future of Engineering

Join the waitlist

Agent Ops opens from this waitlist: on your own, in a live group, or with your team. The other courses are interest lists and open only if enough people ask. The first live cohort is planned for mid-January 2027, and you'll get the exact dates by email.