Chat deleted. So why does the agent still know?
Agent memory makes copies of a single message, and “delete chat” often reaches only the chat. A small lab shows where a diagnosis survives deletion and which deletion path removes it.
Agent memory makes copies of a single message, and “delete chat” often reaches only the chat. A small lab shows where a diagnosis survives deletion and which deletion path removes it.
A grader is a system in its own right, and it needs testing. I had deliberate errors planted in correct release notes and set eleven graders on them. In the end, even the label was wrong.
Three evenings, three tools - and then what? How nine small apps for our volunteer unit became one, why a green health check noticed nothing, and what I do when I'm on vacation.
An LLM feature depends on five things: output schema, prompt, model, tools and eval set. If you only version the prompt, you don't know what you deployed.
The Agent Contract is the intended state, the audit-proof log is the proof of what actually happened. Without the second, the first is just a claim. What DORA, GoBD and the upcoming MaRisk demand right now.
An eval score that only measures quality rewards brute force: more sampling, a bigger model, more retries. Why the right basis for a decision is a cost-quality-latency frontier, a curve instead of a point, and how to pick the operating point in dollars instead of percent.
As soon as an agent does something with consequences, chat is rarely the best interface. A look at the editor, diff, approval panel and timeline - and when a text box is still the right tool.
Permissions, tenant isolation, overly broad tools, sloppy context access: why an agent's reach determines the damage - and how to contain it.
Autonomy is a dial, not a switch. Why most production AI systems get better when the model only sits at clearly bounded decision points - instead of steering an open agent loop.
Green eval in the PR, red numbers in production two weeks later - without a new deploy. Why AI features need their own gates between merge and 100% traffic.
Without traces, model and prompt versions, tool spans and cost metrics, every incident stays guesswork. A guide for teams running AI features in production.
Classic PKM fails at maintenance. An LLM-maintained wiki fixes that - and turns thinking tools from a book on the shelf into a skill on demand.
Volunteer organizations need their own software, but they could never afford it. How AI is changing that right now, and why deployment knowledge matters more than knowing how to code.
Human in the loop isn't a weakness, it's security architecture. Why pause/resume for critical agent actions may be the most important feature of all.
How I used AI to build an offline PWA for drone operations in civil protection - and why domain knowledge matters more than prompts.
How my blog applies the same AI principles I described in my last three posts - with automated podcast generation and AI-assisted copy editing.
A tool call can move money and change accounts. Why idempotency, auth, and audit in agent architectures aren't a nice extra but a hard requirement.
... and start being software - with all the testing obligations that come with it
How I turn a GitHub issue into an implementation plan that's testable and observable