Practical agent systems for people who care about craft.
agent runtimes · workflow tools · evals · human-in-the-loop interfaces
Agent Club is a home for practical agent systems: small autonomous workers, command-line copilots, workflow bots, evaluation rigs, and the glue code that helps people ship with them.
We like agents that are useful under pressure: observable, reversible, testable, and built with enough taste that humans actually want them in the loop.
| Track | Focus | Status |
|---|---|---|
| Agent runtimes | Lightweight agent loops, tool routing, memory boundaries, and task state | Opening soon |
| Workflow kits | Reusable patterns for reviews, release work, research, docs, and operations | Opening soon |
| Evaluation | Fixtures, traces, scoring helpers, and regression checks for agent behavior | Opening soon |
| Interface experiments | Chat, CLI, desktop, browser, and hybrid control surfaces | Opening soon |
- Ship agents, not demos. The goal is durable workflow value, not novelty.
- Keep humans in command. Autonomy is useful when review, rollback, and consent are designed in.
- Make behavior inspectable. Logs, traces, plans, and evals are part of the product.
- Prefer small sharp tools. Composable agents beat giant opaque systems.
- Respect private context. Secrets, user data, and local state deserve hard boundaries.
The organization is just getting warmed up. Public repositories will appear here as they become useful enough to share.
If you are building something in this space, start a discussion around:
- agent architecture and tool interfaces
- eval design and reliability checks
- local-first workflows and privacy boundaries
- UI patterns for supervising agent work
- weird but useful automations that survived contact with reality
We use short names that say what the thing does:
agent-loopworkflow-kiteval-labtoolbenchdesktop-agents
Agent Club is built for people who enjoy the craft of making AI systems feel less magical and more trustworthy.
Pull up a chair. Bring traces.