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 April 3 - April 9, 2026

What I Worked On

Autonomous mode was last week's headline. This week was about teaching the agent your way of working, not just letting it loose.

The main build is an extensibility system: skills, commands, and agents, with multi-tenant CRUD so an organization can define its own. A skill packages knowledge or a repeatable procedure. A slash command gives you a quick way to invoke one, with autocomplete and a /commands list so you can see what is available. Agents can inherit from each other, so a specialized agent starts from a general one instead of from scratch. This is how the platform stops being generic and starts knowing your codebase, your conventions, and your workflows.

The part I care most about is the trust model. Some skills are just text, but some run scripts, and a script that runs unreviewed is a security hole with a friendly name. So skill script execution goes through an explicit approval and trust model. You decide what is allowed to run. Extensible and safe are supposed to be in tension, and this week was about refusing that tradeoff.

I proved the whole thing by pointing autonomous mode at real work: DevNitro built out the blog and comment system for (an early version of) HowdyChad, on its own. Watching the platform build the place its own build logs would live was a good week.

What the AI Did Well

The skills, commands, and agents CRUD is a wide, multi-tenant feature with UI, API, and providers, and the agent built it cleanly with integration tests for the resolvers. It also did security work I do not want to hand-write: centralizing all company authorization logic into one place with tests, and securing the SignalR hubs and billing APIs so nothing sensitive is reachable without the right checks. And it added download and installation telemetry, plus product-health metrics on the admin dashboard, so I can actually see adoption instead of guessing. While I didn't hand-write the security work, I did review it and approve it, and passed it on to additional agents to review. This week Anthropic announced their new model Mythos, which is so powerful that they weren't releasing it to the public yet because of its potential impact. I'm already very security conscious, but this week reinforced that I need to be even more careful about the code that is written. It matters even more when there are companies that utilize your code as the base for their own products.

What Needed Correction

The visual verification from last week brought a packaging headache. The browser automation uses Playwright, and I did not want to bundle its heavy binaries into every build, but I did need the runtime present when it actually runs. Getting that balance right, present at startup, not bloating the package, took a few passes.

I also built a tool retry system with a retryId, because when the language model needs to retry a tool call, doing it efficiently, without re-sending everything, matters for both speed and cost. And I moved a lot of energy this week into naming the category and defining the messaging around it. DevNitro has been hard to market because it is not a single feature, and the platform is not a single product. It is a new way to run engineering, and that is hard to say in a few words. I have more work ahead of me on this.

What Shipped

DevNitro has a skills, commands, and agents system with multi-tenant CRUD, agent inheritance, slash commands with autocomplete, and a secure approval-and-trust model for script execution. It has centralized company authorization, secured hubs and billing APIs, download and installation telemetry, product-health metrics, and a tool retry system.

The Numbers

Category	Hours
Product / Features	19h 33m
Testing	7h 0m
Content	6h 52m
Docs / Skills	5h 33m
Business	1h 21m
Deploy / Infra	1h 7m
Community	1h 6m
Total	42h 31m

A lighter product week by hours, because a big chunk went into marketing and category work. That is not a detour. A platform nobody understands is just a private tool, and this week was about making sure people can understand it.

An assistant does what you say. A platform learns how you work and does it safely, again and again. Skills, commands, and a trust model are how DevNitro crosses that line, and having it build this very blog was the proof I needed that the line is behind us.

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

April 10, 2026 Build Log: Custom Skills and a Trust Model