AI tooling
2026 · open source · 4 tools
I use coding agents every day and build them at work, and these are the tools I ended up writing around them. They're ordered by where they sit in an agent's run, from what it reads before it starts to checking what it says at the end.
01 · Memory · Across every session at a job
Workbase (opens in a new tab)
A notes folder for each job, with the company's repos cloned inside it. Its manual tells the agent where decisions, wins, acronyms and fixes go, and where to look before answering. On my own work notes, it took a test of ten real lookups and captures from 4 passing to 10.
Template · setup script · evals · MIT
02 · Context · Before the agent starts
Agent skills
Skills are markdown files that tell a coding agent when to do something and how. I keep the shareable ones in one public repo, and a small command links them into every repo I work in, so a fix made in one place reaches all of them.
24 skills · Claude Code and Codex · MIT
03 · Evaluation · While a skill is being written
Skillsmith
Skillsmith drafts a skill from the repo, then runs the same tasks with and without it. The judges don't know which result is which, and the skill is only kept if it wins on cases it wasn't tuned on.
Python standard library · part of the skills repo
04 · Runtime checks · After the agent calls its tools
Groundplane
A Python library that checks an agent's structured answer against the tool results it recorded. If the model names the wrong winner or gets a total wrong, it raises an error that shows what the data actually said.
6 checks · 169 tests · MIT
// NEXT
Compoze
Document-grounded assistants, a company I ran alone and sold to a client