Guardrails: A call for taxonomy
Everyone's talking about guardrails in agentic systems, but what does that actually mean? A look at a few different categories of guardrails I've come across and why each are important.
FILTER//TAG 10 MATCHES
10 entries · All tags
Everyone's talking about guardrails in agentic systems, but what does that actually mean? A look at a few different categories of guardrails I've come across and why each are important.
How I decide what an agentic system hands to software, to a model, or to a human: break the work into functions, then classify each by the judgement it needs and the risk you can tolerate.
Sandboxing local agents with macOS Seatbelt instead of containers or VMs: filesystem roots, tool allow-lists, and the quirks of Xcode, Homebrew, and links.
Taking the ideas from the series and turning them into a runnable experiment in agentic engineering.
Putting the pieces together: how I’m using agentic systems in my daily workflow and the patterns that have stuck along the way.
A context parameter seemed like a small addition to my MCP tools. Then I watched the outer harness use it to steer an investigation, again and again.
I started by asking whether autoregressive LLMs could ever reach AGI. The better questions turned out to be what AGI means and where we draw the boundary around the system expected to achieve it.
Keep the workflow, swap the coding harness: using MCP as a portable boundary around bounded, durable agent applications
Runtime QA is judgment from end to end, so let the agent own the investigative loop. Separate its judgment from reviewable mechanics, stop for the decisions that carry consequences, and leave the last word with a person.
The most important question in AI architecture isn’t which model you use. It’s who owns the control loop. That choice affects cost, reliability, and how easily you can swap providers. In this article, I look at three categories of workflows and explore why the nature of the problem should determine the harness, and why code, rather than the model, should often drive the loop.