Notes on calm tools, context, and multi-model work.
Essays, product updates, and field notes on building software with multiple AI models sharing the same memory.
How to run AI coding agents on your repo and stay in the loop
Most ways to run an AI agent on your repo hand you back a diff and take the reasoning with them. This post reframes the problem around three black-box failure modes — you can't trust what you can't see, the reasoning evaporates, and you can't course-correct — then walks through running agents on cloud VMs against your real GitHub repo. Watch the run stream into a live terminal, steer the agent mid-session, and have it write back to your workunit so the next run starts ahead instead of from zero.
Context engineering for solo developers
"Context engineering" has quietly become a real discipline, but almost everything written about it assumes a platform team, a budget, and a fleet of agents in production. Solo developers and tiny teams are the people who feel the context problem most acutely — you are the only one holding the project in your head — and nobody writes for you. This is that piece: a practical take on keeping the context your AI assistants need accurate and close at hand when the context engineering team is just you.
What are context atoms (and why your AI agents need them)
A single markdown blob of "AI context" feels tidy until two agents quietly disagree about a decision you thought you'd settled, because the reasoning was buried three paragraphs down or never written at all. Context atoms are the shape I landed on instead: small typed records — decision, insight, question, attempt, progress — each with an importance level and the ability to supersede an earlier one. This post explains what they are, why typing and ranking memory matters once more than one agent is involved, and how a supersede chain keeps an old decision from silently coming back to life.
A multi-agent coding workflow that actually holds up in practice
One agent is fine right up until it isn't. The moment you want a second model on the job, the obvious approaches start costing you: a master orchestrator prompt that nobody can debug, or copy-pasting context between terminals until your afternoon is gone. This is a concrete account of a different setup — three AI CLIs working the same project through a shared workspace they can all read from and write to, no orchestrator and no master prompt. What each model ends up owning, how they hand off, and where it breaks down.
How to give your AI agents persistent memory (and why most solutions only solve part of it)
Agent amnesia isn't one problem — it's three. The context window fills up and old turns fall off the end. You switch models and start from zero. And the reasoning behind a decision evaporates the moment the session closes. Vector stores, memory files, and hosted chat memory each patch one or two of these and quietly leave the third. This post pulls the three apart, shows where each popular fix lands, and argues that the missing piece is memory scoped to the work, not to the model or the chat.
How to share context between Claude, ChatGPT, and Gemini without losing your mind
You can move chat history between models a dozen ways — exports, extensions, copy-paste. None of them move the thing that actually matters: what the agent figured out while doing the work. The constraints it discovered, the dead ends it already ruled out, the decision it made an hour ago. This is a walk through why transcript-shuffling never fixes the real problem, and what a shared, model-agnostic context layer looks like in practice.
One workspace, many models: why I built Workunit
In late summer 2025 I was bouncing between Claude, Codex, and Gemini across a handful of small projects, keeping a pile of markdown files in sync by hand so each one could pick up where the last left off. The files became their own mess faster than the work did. This is the origin story of Workunit — the path from that mess to a project management tool where humans and AI models are first-class collaborators sharing one workspace, and the specific decisions (workunits, context atoms, an MCP server) that came out of trying to fix my own day.