Posts
herdr vs cmux, after two months on each
Two agent multiplexers compared after two months with each: where the agents run, how plugins work, remote use, and what cmux does better. read more →
Reviewing code you didn't write
Reviewing code is a separate skill from writing it. How to find the issue that matters, especially when an AI produced the change. read more →
Where I run my coding agents, and why it depends on the box
How I limit coding-agent blast radius: disposable Dell micros for unattended work, with permissions kept on around valuable credentials. read more →
Designing MCP tools for agents
MCP tools advertise contracts to the model. Keep related filters typed and shape safe data in code; load large catalogs progressively. read more →
Why local models get stuck repeating JSON
A faster GPU turned 'too slow' into 'broken': why grammar-constrained JSON extraction sends local vision models into repetition loops, and what helps. read more →
Pydantic AI structured outputs and evals on Bedrock
How I stage model trust on Bedrock: structured outputs to lock the shape, Pydantic Evals for content, and a calibrated LLM judge for the open text. read more →
Building and securing MCP servers with FastMCP
Build a FastMCP 3 server for company data. Add group-based access and an audit trail, then test the complete example offline. read more →
Skills or MCP servers: when you need a server
Skills can drive any CLI, but your laptop's credentials aren't an access model. Once data is shared and access has to be governed, you need an MCP server. read more →
Anchoring specs to code with ast-grep
I use ast-grep rules to keep living specs tied to the code they describe, even when files move and line numbers change. read more →
Breaking In
The entry-level market narrowed and AI writes most of the code now. A principal engineer on whether it's still worth starting in software in 2026. read more →