Practical notes on building AI agents that work in production, with diagrams and sources.
Tell us about one process. We'll show you how an agent could handle it, and what to measure to confirm it's working.
Security · 6 min read
No single product secures an agent. A reference design with five layers, and where tools from Fortinet, Palo Alto Networks, CrowdStrike and the open-source eval projects fit.
Governance · 4 min read
How much control you hand to an agent is a design choice, not a switch. Five levels, from operator to observer, told as a walk with a small robot.
Architecture · 6 min read
OpenAPI (Swagger) describes an API for developers. MCP describes tools for a model. They are layers, not rivals, and the choice shows up in discovery, security and long tasks.
Architecture · 6 min read
Most architecture diagrams mix levels and hide what the arrows mean. C4 gives agent projects four zoom levels. Use the tabs to walk from the whole system down to the code.
Governance · 5 min read
When an agent starts giving wrong answers or burning budget at 3 a.m., the person on call needs steps, not theory. How to write a runbook for it.
Governance · 4 min read
Models and providers change every few months, and the reason behind a choice is gone sooner than the code. A short record per decision keeps it.
Architecture · 6 min read
Models are only one layer. A walk through the request path, and the shared services that keep agents reliable.
Measurement · 6 min read
A token bill says what you spent, not what you got. How to measure agents so finance and engineering agree.
AI OS · 6 min read
The phrase is used loosely. A clearer meaning: the infrastructure layer that enforces the guarantees agents cannot give themselves.
Orchestration · 6 min read
From one agent to many changes the problem from prompts to systems design. Five patterns, two protocols and what breaks first.
Models · 6 min read
There is no single best model. How to match tiers to tasks, when routing pays off, and why voice is its own problem.