AgentRQ multi-agent task board with MCP-driven scheduling and event workflows

What they did

They set up AgentRQ as a shared workspace where human operators and AI agents collaborate on the same task board. Using the Model Context Protocol, agents like Claude and Codex connect to the workspace and can see its state: they pull their assigned tasks, move cards between Not Started, Ongoing, Blocked and Done, request permission for sensitive actions, and post questions back to the human. The human breaks a goal into tasks, assigns each to an agent, and watches the board update in real time as cards move. They used task scheduling to give recurring work a cadence (every 15 minutes, hourly, daily, weekly or custom days), letting a background poller spawn runs when due instead of keeping an agent awake. They wired Events — named signals like qa_passed, deploy_finished or blog_published — so that a finished task fires a signal that automatically creates a new task in another workspace, and arranged those events and workspaces on a Workflow graph by dragging connections rather than writing YAML. For long autonomous stretches they flipped YOLO mode on a task so every permission request was auto-allowed and still recorded in history, then turned it off to restore approvals. They audited runs through the History tab's lane-grouped timeline of Input, Agent and Tools calls, searching for specific calls and inspecting what ran and what it returned, and reviewed the per-task timeline showing time spent waiting on the human versus blocked.

How it went

The board ran end to end: five tasks were picked up by five agents, cards advanced from Not Started to Ongoing, one agent asked a question and its card turned Blocked until the human clicked Allow, then every card landed in Done. Scheduled tasks fired on their cadence, event chains propagated from qa_passed through deploy_finished into release notes, and YOLO mode let a test suite run to 48 passing tests without a single approval wait. Tool-call history and timelines gave a clean audit trail of what each agent did and where time went.

Worth knowing

Requires installing the AgentRQ Chrome extension or platform and connecting each agent through MCP, so setup is more than a single config file. YOLO mode removes the human approval gate entirely, which is risky on tasks that touch production or credentials. Event and workflow wiring depends on consistent event naming across workspaces, and the README does not detail self-hosting, auth or data retention. Agent support varies by client, and scheduling relies on the platform's poller staying up.

Wiring

Install the AgentRQ Chrome extension from the Web Store and open the workspace. In each agent client (Claude, Codex), add the AgentRQ MCP server so the agent can read and write workspace state: list tasks, claim assigned tasks, update status, request permissions and post messages. Create a workspace and its task board, then create tasks and assign each to an agent by name. For scheduling, set a launch date or a recurring cadence on a task; the platform's poller ticks every minute and spawns the run when due. For automation, define named events (qa_passed, deploy_finished, blog_published) and attach them to tasks as completion emitters; subscribe a target workspace to an event so it receives a new task when the event fires. Arrange events and workspaces on the Workflow graph by dragging a workspace onto an event to subscribe it, or an event onto a workspace to emit it on completion. Toggle YOLO per task or per workspace to auto-allow permission requests while still logging them, and review runs in the task's History tab and timeline.

Try it yourself

Set up AgentRQ for this repo: install the AgentRQ Chrome extension, connect a Claude and a Codex agent over MCP to a new workspace, and create a task board with columns Not Started, Ongoing, Blocked and Done. Add three tasks: one that runs the test suite, one that lints and fixes style, and one that drafts release notes from recent commits. Assign the first two to Codex and the release notes to Claude. Give the test task a daily schedule, and wire an event called tests_passed that fires when the test task completes and creates the release-notes task automatically. Leave YOLO mode off so I can approve shell commands, and show me the tool-call history for the first run.