Story

agx is a kanban board where each card is a task that AI agents actually execute.

    agx new "Add rate limiting to the API"

The technical problems this solves:

The naive approach to agent persistence is replaying conversation history. It works until it doesn't:

1. Prompt blowup. 50 iterations in, you're stuffing 100k tokens just to resume. Costs explode. Context windows overflow.

2. Tangled concerns. State, execution, and orchestration mixed together. Crash mid-task? Good luck figuring out where you were.

3. Black box execution. No way to inspect what the agent decided or why it's stuck.

agx uses clean separation instead:

- Control plane (PostgreSQL + pg-boss): task state, stage transitions, job queue

- Data plane (CLI + providers): actual execution, isolated per task

- Artifact storage (filesystem): prompts, outputs, decisions as readable files

Agents checkpoint after every iteration. Resuming loads state from the database, not by replaying chat. A 100-iteration task resumes at the same cost as a 5-iteration one.

What you get: - Constant-cost resume, no context stuffing

- Crash recovery: agent wakes up exactly where it left off

- Full observability: query the DB, read the files, tail the logs

- Provider agnostic: Claude Code, Gemini, Ollama all work

Everything runs locally. PostgreSQL auto-starts via Docker. The dashboard is bundled with the CLI.