An agent space with a place for the next step.

An agent space starts with the work, not the conversation

A chat is useful for deciding what to do next. A Project gives the work a durable home: source files, instructions, Sessions, and saved results. Put the current brief and accepted deliverables there so a later task can read the evidence instead of relying on a remembered conversation.

For a fictional client engagement, keep brief.md, the approved requirements, a draft report, and handoff.md together. State which file is authoritative, who reviews the result, and what is still a proposal. A shared space helps only when the next participant can distinguish current decisions from unfinished suggestions.

Give each assistant a clear role

Connect each client before asking it to work

Use the hosted MCP endpoint https://workspace.clankercloud.ai/mcp and the client’s supported OAuth flow. Start with a read-only workspace_status check. Then name the Project, allowed inputs, required output, and actions the user has authorized. Read, Edit, and Configure and run are separate permissions.

A Project name in a prompt selects task context; it does not narrow the OAuth grant to that Project. The grant follows the connected user’s existing Workspace role. Another assistant does not acquire access merely because its name appears in a file. Hermes and OpenClaw setup follows their current client documentation; verify your installed version and connection before relying on it.

Sources: Clanker: connect an external agent and review permissions

Hand off an artifact and a decision

Ask the first assistant to save what changed, the exact output filenames, unresolved questions, and the next permitted step. The next authorized assistant reads that packet in a fresh task. For example, ChatGPT can save a proposal summary and Claude can review it against the same approved criteria.

Saved Session messages are records. Through the external MCP connection, sending a message to another Session saves a prompt; it does not by itself start an autonomous Agent or run a second model. Decide explicitly who should act next and verify that the requested work actually ran. Keep private reasoning and credentials out of handoff files.

Make the shared space useful to its owner

Open the actual saved output in Workspace. Check important statements against the source, confirm revisions, and keep missing information visible. Use a reusable Agent when the method should repeat; use a new task when the brief changes. An agent space is valuable when a person can inspect what happened and continue the work.

Projects are saved context, not a promise of a continuously running computer. Choose cloud compute only when a task needs execution, and configure recurring work separately. Review sharing deliberately before inviting people or publishing an App.

Give your assistant its first Workspace task.