Don't Trust the Model. Build the Checkpoint.
A pattern for reliable agent workflows: move the truth into executable checks, scripts, and oracles the model has to pass through.
Thoughts on software engineering, product development, and building things right.
A pattern for reliable agent workflows: move the truth into executable checks, scripts, and oracles the model has to pass through.
Every agent framework implements retrospective memory. None of them implement prospective memory. Cognitive science has known about this gap for 35 years.
I ran Sonnet 4.6 through five different coding CLIs, three times each, and traced every API call. Same model, wildly different context profiles.
Agent memory only looks backward. I added a forward-looking type and benchmarked it against cron scheduling. Projection scored 4.80. Smart cron scored 2.59.
I pointed GLM-5, Kimi K2.5, MiniMax M2.5, and Sonnet 4.6 at the same task. The differences showed up in planning, not raw capability.
Claude used 23K tokens to fix a one-line bug. Gemini used 350K. I built a tool to find out why.
Coding agents write the code, pick the next feature, and fix the bugs. You run the tests and say 'looks good.' Is that still engineering?
"Full stack" sets expectations nobody can meet. The real gap isn't technical range, it's everything around the code.
Using docker:dind plus docker compose in GitLab CI makes full integration tests feel local and repeatable.
Ambition gets healthier when you focus on the work you want to do, not the title you want to carry.
The hard part of building software has never been the building, but understanding why, for whom, and what happens if you get it wrong.
When a client has been burned by a previous team, no amount of technical skill will save you. You need rhythm.
Agile is fun. Failing early is great. But failing all the time isn't.
Let's set up a development environment for a Python and Java service in Docker, featuring code watching, Gitlab CI, and other good stuff.
REST is a heavily oversimplified and massively misunderstood (but very widely used) concept for designing APIs. Here's a better approach.
Being able to program is one of the few skills that can make you be seen as less than what you are.
A satirical commentary on Java's date/time libraries.
Improving the efficiency of IoT systems development through better tooling.
The journey to improve software development by orders of magnitude.
Scrum is great, but it has some disadvantages. Let's discuss a potentially even better way of doing Scrum!
My experiences after a few years of applying SCRUM for a software development process.
We're still developing!
Don't let new and shiny tech lure you into finding a problem for it. First identify the problem, then find the tech to solve it.
How I've found a positive approach to management.
Speaking up, disagreeing with intellect instead of emotions, and don't be an ass.
Smart things are stupid. Smart systems are smart.
What is LoRa, and how can you use it?
Let's be lazy and deploy simple docker-based architectures directly through docker compose, using only free tooling.