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 17 - April 23, 2026
What I Worked On
The whole idea of a blueprint is that the AI inherits your architecture instead of guessing at it. But architecture changes. If a project can only inherit a blueprint once, at creation, it drifts away from that architecture the moment the blueprint improves. So this week I built a blueprint upgrade system. An existing project can pull forward changes from a newer blueprint, with a robust token replacement engine so your project-specific names survive, upgradeIgnore rules so your custom edits are not stomped, custom folder exclusions and auto-delete for files that should go away, and support for binary files and older manifest formats so nothing breaks on the way up.
The test that mattered: I upgraded a real production website by pulling its blueprint forward. It worked. That is the loop I have wanted from the start. Improve the blueprint, and every project built on it can come along, safely.
The other theme was making parallel autonomous work safe. When several workers run at once, they cannot share state by accident. So I added true workspace isolation to the worker loop, isolated the conversation context with AsyncLocal so one agent's thread cannot bleed into another's, and made the max number of concurrent workers dynamic instead of fixed.
What the AI Did Well
The agent did a large refactor toward vendor-neutral abstractions: a transactional email abstraction with the concrete provider behind it, pluggable email-marketing providers, and Azure Storage split into its own package. That is the same provider-independence instinct behind running any model you want, applied to infrastructure. It also modernized the test suite into focused classes on .NET 10, replaced brittle sleeps with proper draining, scoped memories to team and project, and added API rate limiting and security hardening.
What Needed Correction
The test cleanup addressed past work done by agents that I had not properly reviewed. Verification is part of Intelligent Engineering, and time was just spent cleaning up test code that wasn't properly verified.
Yet, this verification requirement became more real this week for another reason. I pointed an autonomous run at a set of issues and it hallucinated child issues and updated the wrong ones. Because issues live in a database, a confused agent can do real damage, not just waste a turn. That is a different class of risk than a bad code suggestion. My takeaway was that the issue system needs guardrails: at minimum, refuse to link issues across projects, and consider versioning so a stray edit can be caught and undone. An agent with write access to your backlog has to be governed like one, not trusted like a saint. Having a model update your code that is version controlled is one thing. Having it do work in your database (even via a tool working with your data) is another.
I also got real user feedback that turned into a bug list: running in debug mode on a different port caused CORS problems, the model was not consistent about moving finished issues into review, and the web conversation view did not refresh after a dropped connection until you toggled it live. Small things that add up to whether the thing feels trustworthy.
What Shipped
DevNitro has a blueprint upgrade system with token replacement, ignore rules, exclusions, and binary and legacy support, proven by upgrading a live site. Parallel autonomous work is isolated per workspace and per conversation, with dynamic worker concurrency. The framework moved to vendor-neutral email and storage abstractions. Memories are team and project scoped, and the APIs have rate limiting and hardening.
The Numbers
| Category | Hours |
|---|---|
| Product / Features | 36h 9m |
| Content | 4h 5m |
| Deploy / Infra | 2h 1m |
| Community | 1h 2m |
| Business | 0h 29m |
| Testing | 0h 13m |
| Total | 43h 59m |
Nine of those hours were labeled plain "bug fixes," and I am fine with that. The week I let an agent edit my backlog and watched it get creative is the week guardrails stopped being theoretical.
Inheriting architecture continuously, safely, as it improves, is a blueprint. That distinction is the heart of Intelligent Engineering, and this week it went from a word to a working upgrade path. Verification is another piece of Intelligent Engineering, and this week it went from a concept to a hard lesson. Both are necessary for a system you can trust, support, and scale.
Curious how this site was built?
Watch the 5-minute walkthrough at devnitro.com →
0 Comments
No comments yet. Be the first to share your thoughts!
Leave a comment