The default for AI-assisted development is one of two failure modes.
Either you're babysitting the agent line by line — approving each diff, re-ex...
For further actions, you may consider blocking this person and/or reporting abuse
The three-gate model is a good way to keep human judgment where it matters. I would add one more explicit artifact between HOW and GO: the rollback or failure policy. Agents are getting good at implementation, but teams still need to decide what happens if a migration half-succeeds, an assumption is wrong, or a generated change touches a shared contract.
I agree completely. All work is delivered via stacked PRs via git worktrees. Nothing lands on main until the phase has gone through AI-agent cross-platform internal review before PR is opened and we also support PR-review services like CodeRabbit as part of the delivery. The stacked PR implementation of a phase makes sure each PR is 300-500 net lines of code (i.e. human reviewable a vertically sliced).
That stacked-PR shape is exactly the part that makes the model usable instead of just fast. Small vertical slices plus a second review layer gives humans something they can actually inspect. I like the 300-500 net-line target because it turns "AI did it" back into normal engineering review.
See it in action.
I just converted my coding metrics site (codingstats.vercel.app) completely from legacy svelte4 idioms to idiomatic svelte5 runes syntax. A 4-hour long refactor in Son of Anton
cookmode. The primary agent moves through an entire phase delivery end-to-end, ticket-by-ticket until the phase is complete ... no baby-sitting the AI after/soa plan(the WHAT) and/soa decompose(the HOW) decisions have been locked in - yet the work is delivered in a way I retain full control and auditability on what was delivered and what will be/soa closeout(the DONE).github.com/cesarnml/coding-stats/p...
8 stacked PRs ... all fall between 300-500 net code line changes. Each step "human" reviewable vs vibe-coding the same change would have been a ghastly 2000+ line PR. At that point, we lose the developer connection to the codebase.
Thank you for the comments and the support.
Phase delivery retrospective by the agent: Phase 6: Svelte 5 Runes Migration Restrospective
That stacked phase shape is the part I like: lock the what and how first, then let the agent execute tickets while the audit trail stays readable.
The control point is not watching every keystroke. It is making sure each phase has a clear decision boundary, testable output, and a way to stop before the next stack of changes starts.
Historically, software engineering rewarded people who could build. AI increasingly rewards people who can decide. That's a subtle but important shift.
I am glat to see good post
The three gate model is a great way to keep human judgment focused where it matters most.
One addition I would make is an explicit rollback / failure plan between HOW and GO.
Agents are getting very good at implementation, but teams still need to decide what happens if a migration only partially succeeds, an assumption turns out to be wrong, or a generated change affects a shared contract.
Defining recovery paths, rollback triggers, and ownership upfront helps ensure that execution is not only fast, but also safe and reversible.
I wish you write good post more than before
Great framework on the three gates. The Scope Lock phase is where I see the most friction in practice — agents tend to skip straight to implementation before scope is settled, and that execution drift costs more time than it saves.
I put together a small plugin to address exactly this: Brainstorm-Mode (mehmetcanfarsak/Brainstorm-Mode on GitHub). It uses PreToolUse hooks to block premature tool calls during the ideation phase, keeping the agent focused on scoping and brainstorming before it jumps into code. Three modes (divergent, actionable, academic) help match the brainstorming style to the kind of project you are scoping.
Brainstorm-Mode -- very interesting ...
Yes Gate 01 (The What) and Gate 02 (The How) also prevent agent execution. It's where the developer engages with the AI-agent as a thought companion. The only deliverables are explicit Gate 01 delivers the phase product plan. Gate 02 delivers the phase implementation plan and the ticket decomposition.
does this post still have loading to the end? i dont see
.at the end. or does this new article method?I'm not sure I follow, friend. Let me know if something seems amiss.
The three-gate framing is strong because it keeps human judgment where it matters: intent, design tradeoffs, and final acceptance. The danger with agents is not letting them write code. It is letting them silently decide what problem they are solving. Making the decision points explicit fixes a lot of that.