OpenRig

The daemon in the OpenRig architecture serves as the infrastructure layer, persisting chat room records, maintaining inbox/outbox status, logging seat event streams, and providing real-time data for the dashboard and views.

rig ask does not call another AI; it only runs a plain-text search on the saved transcript text, but offers the option to wake a saved session for direct questioning, which runs a real Agent and consumes tokens.

OpenRig's transcript rolls and saves copies of each screen like a dashcam, but only retains recent history and discards the earliest portions, and is not a complete recording.

OpenRig's communication philosophy can be distilled as: use send to ask, use queue to record accountability, use capture to observe, and use transcript to search.

The RIG commands used by Agents in OpenRig are exactly the same as those used by human operators, and both run them directly in their respective terminals.

The key to choosing a communication method in OpenRig depends on two questions: who needs to hear this message, and whether it needs to leave a persistent record.

RigSend is a point-to-point communication method that inputs messages directly into a specific Agent's terminal, offering two modes: waiting for the target to be idle and checking the screen after sending to confirm receipt.

In OpenRig, Delivered only means the message has been input, not that it has been read.

queue is the team's official memo board introduced in OpenRig from Episode 6, serving as an accountability anchor where each item requires a specific seat to explicitly claim or reject it.

RigBroadcast is used to notify multiple Agents at once; forgetting to specify a particular rig will spread the message to all sessions on every rig including the kernel, and using --pod alone will match the dev pod on every rig.

Each rig has a Chat Room persistently maintained by a daemon, which serves awareness purposes and cannot replace the queue.

Bad ideas can spread among agents like thought viruses, and the author draws on epidemiology to perform 'contact tracing' and 'inoculate agents with memetic vaccines' to stop the spread

OpenRig's core goal is to enable existing agent frameworks such as Claude Code and Codex to work together collaboratively, making it a 'framework of frameworks'

The author runs six OpenRig instances — three on Mac Minis at home and three deployed on VPS servers at OVH Cloud — communicating with each other over a Tailscale network

One agent can inject keystrokes directly into another agent's terminal using tmux's send-keys command, thereby increasing collaboration speed by approximately 50x

OpenRig's most critical abstraction is the 'seat,' which defines roles, skills, and model settings, provides a stable address (e.g., builder@workshop), and enables knowledge inheritance

OpenRig organizes work into a hierarchical structure of projects, tasks, and slices, where a task typically corresponds to a complete release version, and a wave diagram shows which slices can be executed in parallel or must be executed sequentially

Bureaucracy is a typical failure mode of agent swarms: checks are continuously added to prevent errors, until 'completing the process' itself becomes the goal everyone is busy with, and the rules become more important than the matters they serve

Scope creep is a failure mode of agent swarms: each local decision seems reasonable, but the cumulative result is something nobody wanted, and it happens faster the more agents there are

OpenRig has a feature called 'refocus' that tracks the chain of intent at each level and sends that intent chain to agents after compression or at intervals to check whether they are aligned with the bigger picture