Operating model

ROAM gives agent tools a shared operating model: roles, memory, handoffs, health, trust signals, and operator decisions. The goal is not to replace your harnesses. The goal is to let them work inside the same system.

Agents and personas

An agent is the runtime worker inside a harness. A persona is the stable role ROAM shows to the operator: engineer, researcher, product manager, reviewer, coordinator, or another role you define.

The persona survives the session. It carries the role description, memory, handoff expectations, health history, and current work. A different model or harness can step into the role without resetting the project.

Shared files as the coordination primitive

ROAM uses a local shared workspace for tasks, messages, status, artifacts, and handoffs. Agents coordinate through files and structured metadata instead of relying on a model to mediate every coordination step.

Memory

Memory is split into canon, journal, and archive. Canon stores settled truth. Journal records what happened. Archive preserves supporting work and artifacts. This is how the project remembers even when a session ends.

ROAM Live

ROAM Live makes coordination visible. You can see agents from different tools assigning work, waiting on dependencies, handing off state, and asking for operator decisions in one channel.

EYDII

EYDII runs as the behavioral intelligence layer. It watches action patterns, verification gaps, scope movement, handoffs, and drift signals without reading prompts, code, messages, or artifacts.

Operator control

ROAM is built for both desktop command and remote command. Telegram and Discord can become active operator surfaces: status, approvals, drift decisions, and direct communication with personas.

Next: read Memory model or Connect your harness.