Two GPU rigs, several agents, and no way to see who was doing what.

It is not a small problem. An agent that cannot see the others duplicates their work, and the duplicate does not look like a mistake in a terminal window; it looks like progress. The human in the middle cannot tell the two apart either, which is how a morning disappears into two versions of the same fix.

Panopticon is the board the agents check.

What is in it

• A dashboard: KPIs, fourteen-day throughput, a per-project table, live agent status, recent completions, refreshed every five seconds.
• A kanban: an inbox of open work plus a board per project, with drag-and-drop, claim badges, delete, and a task detail panel.
• An MCP server with twelve tools — list and create projects, list and create tasks, update status, claim, delete, plus agent status, skills, LLM endpoints, forges and profiles. That last set is what makes it useful: an agent discovers the fleet’s capabilities from the same endpoint it claims work from.
• Agent check-ins: the Jido agents on supra and turbo post their status, so the fleet is visible in one place instead of three terminals.
• Forgejo integration: a finished task can open a pull request against the git server the work actually lives on, rather than a pasted diff in a chat.
• Emporoman integration for the ISATEK workspace, because half my work starts as an ERP task.

The claim is the part that matters. Everything else is a view onto it. A claimed task is a task nobody else touches, and that rule removed most of the wasted mornings.

Deliberately out of scope

Multi-tenant SaaS, websocket channels for push updates, agent-to-agent protocols, and a distributed queue across machines. Single node, SQLite, five-second refresh is enough for two rigs and one person, and every one of those omissions is a feature I did not have to debug.

What it works against

The idea that agentic work is a sequence of prompts. It is a small organisation, with collisions, handovers and priorities, and it needs what any small organisation needs: a place where the work is written down and who is on it is visible.

It also works against my own tendency to keep the plan in my head. Once a task is on the board with a claim, I cannot pretend two agents are not about to touch the same file.

The honest part

Internal tooling. It gets attention when coordination breaks down, not when it works, so parts of it still look unfinished. A dashboard nobody reads in a good week does not get redesigned in a good week. If it were for sale it would look different; it is not, and that is the right trade-off for a tool used by two machines and one person.