A note about this build log entry: I published this after the week it covers. DevNitro reconstructed a draft from work-related entries in my journal, Git commit history, and time tracking. I reviewed, edited, and approved the final post. The story reflects what I knew and built at the time.

From the Backlog to the Build Log

Week of May 15 - May 21, 2026

What I Worked On

If you want several agents working at once, the first question is not speed. It is: how do you keep them from wrecking each other's work? The now well known answer to this problem is Git worktrees.

A worktree lets you have multiple working copies of the same repository, each on its own branch, side by side. So I finally made concurrent worktrees the isolation layer for parallel agents. Each agent gets its own worktree with its own workspace context, tracked through AsyncLocal so one agent's state cannot leak into another's. I refactored the git tools down to two clean primitives, git_start and git_commit, taught them to handle a parent branch and sync changes across worktrees, and, most importantly, protected the main and master branches so an autonomous agent physically cannot commit to them if the team's git strategy forbids it. This is visible parallel work with real fences around it, not just several agents sharing one checkout and hoping.

I also kept extending the human-readable references from week before last: sessions now get per-company numbers like S123, conversations got user-facing refs, and I cleaned legacy ID fields out of config so old identifiers cannot pollute the new slug-based routing. Small, but it is the connective tissue that makes the whole thing feel coherent.

And really, a big chunk of the week was content: I spent around twenty hours recording autonomous mode building this HowdyChad blog, trying to get a full build down to about ninety minutes on camera. Dogfooding on video is its own kind of pressure test.

What the AI Did Well

The worktree refactor is exactly the kind of careful plumbing the agent is good at once the design is set: two primitives, parent-branch logic, cross-worktree sync, and integration tests. It also fixed a genuinely nasty bug, a pipe-buffer deadlock in git process execution, where a git command could hang forever if its output filled the buffer and nobody drained it. Those are miserable to find by hand.

What Needed Correction

The first cut of worktrees did not actually work with the tools. I recorded a build and watched the tools ignore which worktree they were supposed to be in, and two running instances of the same app collided because they grabbed the same port. So I sent the agent off overnight to make the tools respect the active worktree and to give each running app instance a unique port. Concurrency exposes every hidden shared resource you own, and a port is a shared resource you forget about until two things want it.

I also fixed a Server-mode bug where automated workers kept showing as active after a run had finished, and a reporting bug where admin date ranges returned the wrong results, which a user had flagged. Small correctness issues, but they are the ones that make people quietly stop trusting a dashboard.

What Shipped

DevNitro runs parallel agents in isolated Git worktrees with AsyncLocal workspace context, protected main and master branches, git_start and git_commit primitives, cross-worktree sync, and unique ports per running app instance. Sessions and conversations have human-readable refs, legacy ID fields are cleaned from config, and a git pipe-buffer deadlock is fixed.

The Numbers

Category Hours
Product / Features 24h 42m
Content 23h 49m
Deploy / Infra 1h 55m
Community 0h 23m
Testing 0h 8m
Total 50h 56m

Almost an even split between product and content this week, because filming autonomous mode building a real site is both. Twenty-five hours went to worktrees, refs, and bug fixes: the unglamorous foundation for running many agents without chaos.

Parallel work is only a feature if it is safe. Isolated worktrees, protected branches, and unique ports are what turn "ten agents at once" from a scary demo into something a team can actually run. That safety is the whole point of Intelligent Engineering.

Curious how this site was built?
Watch the 5-minute walkthrough at devnitro.com →

An AI built this blog (but not this post)

This entire blog system (domain models, database, services, UI, auth, etc.) was built by DevNitro on its own in 95 minutes. I recorded every second.

    We respect your privacy. Unsubscribe at any time.

    Built with Kit