100% of your code will be written by agents.The only question is which level you’re at.

You could, and it would sound right. Ask one about Codex approval modes and there is a good chance you get `--approval-mode`, which is not a flag — the real one is `--ask-for-approval`, and its values are `untrusted`, `on-request` and `never`. Copy the invented one and Codex rejects the command outright — you do not get a wrong permission posture, you get an agent that will not start. What you are paying for is that someone checked, against a named release, and wrote down the version they checked against. It matters more the higher you run: an agent you watch at Level 2 gets corrected in a minute; an agent left unattended for twelve hours on an invented flag is twelve hours of nothing, discovered in the morning. We keep a list of the ones models keep inventing.

Real, and dated. Boris Cherny, who leads Claude Code, told Fortune on 29 January 2026 he had been at 100% for two-plus months and does not make small edits by hand. Anthropic reported that over 80% of the code it merged in May 2026 was written by Claude. Stripe’s agents merge 1,300 PRs a week (February 2026) with no human-written code, every one human-reviewed. The honest caveat is the other half of the research: METR measured experienced developers 19% slower with early-2025 tools, and its 2026 data trends toward zero in a study METR itself is redesigning; Faros’ 2026 telemetry shows incidents per PR up 242% across 22,000 developers where review did not keep pace. The percentage is not the hard part. The level you run it at is.

The discipline. Karpathy’s definition from April 2026: agents are fallible and stochastic, but extremely powerful, and the job is to coordinate them to go faster without sacrificing your quality bar. Humans keep the spec, the review, the evals and stop conditions, the permissions, the taste and the direction. Or, in his words: vibe coding raises the floor; agentic engineering is about extrapolating the ceiling.

That is the most common reason people cancel, so it is worth answering plainly. Start with the Setup Pack — the Harness station — answer nine questions, download the config files for your stack, and drop them into a repo. That is a working harness in about twenty minutes, before you read a single guide — and the guides make sense afterwards because they describe a setup your agents are already running on. If it still is not for you, the 30 days cover it.

Daily use is Level 2: one agent, one task, you watching. Level 3 is where the agent writes and you review; Level 4 is where you write the spec and check the tests twelve hours later. Almost everyone tops out at Level 3 (Dan Shapiro, January 2026). The guides are the moves between rungs — hooks, verification gates, worktrees, overnight runs, factories — checked against the current release of Claude Code, Codex and Cursor.

Dan Shapiro published the Five Levels in January 2026; on this site they read: Level 0 by hand, 1 assisted, 2 paired, 3 code-review manager, 4 spec manager, 5 dark factory. The free scorecard maps its 25 questions onto Levels 1 to 4 — Level 0 does not take the test and Level 5 cannot be verified by one — and the guides are organized as the moves between them.

A YouTube tutorial was true the day it was recorded. These tools change weekly — flags get renamed, defaults move, models get retired — and a video can't be patched. Every guide here is re-checked against current releases and updated as they ship. And it's a curriculum, not a playlist: each guide builds on the one before it, from first install to enterprise rollout.

Each guide takes 15–30 minutes, and you do not read the ladder top to bottom: find your level with the free scorecard and read the one hub above it. The Quick Wins recipes are copy-paste — start there and go deeper when a technique earns it.

Yes. One click in your account settings. No questions asked, no hoops to jump through. Your access continues until the end of your billing period.

Yes — 30 days from your first purchase, for any reason. Email support@developertoolkit.ai and we refund it; we won't argue with you about why. Applies to first-time purchases, as set out in the Terms.

You keep full access until the end of the period you've already paid for — the subscription simply doesn't renew. Your account stays in place, so you can pick up where you left off anytime. Any ebooks you've downloaded are yours to keep.

Yes. Payments are handled by Polar, our merchant of record, which issues an invoice for every charge with the applicable VAT or sales tax for your country. It lands in your inbox right after payment — ready to expense or pass to your accountant.

Unlimited seats for one flat price. Everyone with an email address on your company domain gets full access automatically, and teams spread across domains can add named addresses instead — no per-seat billing, no invites to manage. Need custom onboarding, white-label, or an on-premise option? That's the Enterprise tier.

Our 100+ recipes cover React, Next.js, Vue, Python, Node.js, Go, Rust, and more. And the core of the system — context engineering, rules, MCP, review workflows — doesn't care what language you write. The recipes just make it concrete in yours.