Blog
Notes from production - agentic AI, cloud economics and the unglamorous engineering that ties them together.
Latest · Nov 2026Specialist review agents on a shared message board with Claude Code, Codex and MCP
In Triglav you define your own researcher agents. Mine are a bug hunter, a security reviewer and a planner, and they review plans before any code exists. Four real threads show how they check each other.
Running an LLM agent on untrusted pull requests with Claude Code and GitHub Actions
I built a GitHub Actions workflow where a Claude Code agent writes unit tests for pull requests. Here is how I separated the agent from untrusted code, what broke, and what is still open.
Claude Code and Codex agents on a shared message board: what 12 days of messages show
My coding agents posted 1,648 messages to each other in 12 days. What helped, what did not, and what I want to change in Triglav next.
Triglav is public: a local orchestrator for Claude Code and Codex agents
Triglav is a local orchestrator for Claude Code and Codex agents. After 41 jobs and 96 local review rounds on my own work, the first public version is on GitHub.
Sharing Claude Code skills across repositories with a private plugin marketplace
How we turned an architectural mistake and code review feedback into reusable Claude Code skills, then shared them across six repositories with a private plugin marketplace.
Triglav, part 2: AI agents planning enterprise SSO on AWS Cognito over a shared message board
A design-doc job where Staff, Senior, reviewer and I worked through one message board: questions with recommendations, a proposal, evidence-backed review findings and a verdict, in about forty minutes.
Triglav, part 1: running a team of AI coding agents with Claude Code, Codex and MCP
Why I built a local orchestrator that runs Claude Code and Codex agents as a small consultancy, how the roles work, and what six days of real use showed about where the tokens go.
A streaming LLM chat endpoint with the AI SDK and Bedrock
How our chat endpoint routes a request, stacks system prompts and runs a capped tool loop against Claude on Bedrock, and why one failure was silent.
Hunting a connection leak: from an OOM to a one-line fix in a third-party library
A production OOM that came down to one wrong method call - releaseLock() instead of cancel() - in a GraphQL-over-HTTP SSE executor, plus all the wrong guesses I made getting there.