Blackboard and inbox

Multiple agents do not share the same conversation context. Their collaboration is built on two explicit vehicles: the blackboard carries artifacts, and the inbox carries signals. The two have distinct duties and are used together.

Blackboard

The blackboard is a shared key-value store that any agent can read and write by path. It carries the things that need to be persisted for others to draw on: task ledgers, progress records, interface contracts, review conclusions, and artifacts of all kinds.

Characteristics:

• Many-to-many, persistent — anyone can read and write any key, and the content is saved to disk.
• Version history — every write appends a record, so you can look back at "who wrote what and when" for a given key, and mid-course revisions are traceable.
• Full-text search — keys and content enter the full-text index and are searchable.
• Write means wakeup — when a key is written, agents that declared a dependency on it are woken up immediately (see Wake mechanism).
• External edits are detected — if you edit a blackboard file directly in your editor, that too is detected and recorded in the history.

Scope: keys are usually prefixed with the workspace and direction, so different directions and different workspaces do not overwrite each other. Members sharing the same direction can see each other's keys; whereas Fusion competitors are confined to their own prefixes, cannot see each other, and are thus guaranteed independent implementations.

Inbox

The inbox is a point-to-point message channel. Compared with the blackboard, which carries "artifacts," messages carry the signal that "something happened": requesting help, reporting results, asking questions, giving answers.

• Message types: note (general communication), ask (a request), reply (an answer), wake (a system wake signal, not human mail).
• Threading: a reply can hang below a given message, forming a conversation chain.
• Read tracking: messages record their sent, delivered, and read times; an agent reading its own messages marks them read automatically — but wake signals are the exception, to ensure the end-of-turn check can still see them.

How the two work together

The typical pattern is: first write the artifact to the blackboard, then send a notification message. An agent subscribed to that blackboard key is woken by the write; a recipient that is not subscribed is woken by the message. Separating artifact from signal makes collaboration both documented and promptly driven.

The swarm tools an agent uses

The capabilities above are exposed to the model as native MCP tools. An agent calls them just like any other tool:

Tool	Purpose
swarm_send_message	Send a message to another agent (or the user), optionally specifying a type and reply target.
swarm_list_messages	List recent messages sent to itself (marks non-wake messages read).
swarm_search_messages	Full-text search across all messages.
swarm_list_agents	List all agents (online and historical).
swarm_list_blackboard	List all blackboard keys and their latest version info.
swarm_read_blackboard	Read the current content of a key.
swarm_write_blackboard	Write to / overwrite a key, signed as itself.
swarm_spawn_worker	Dispatch a worker, specifying a role and instructions (see Orchestrator and roles).
swarm_list_roles	View the catalog of roles available for dispatch.
swarm_name_thread	Name the current direction once (triggers isolation on a git project).
swarm_fusion_consult	Start a multi-model research committee consultation.

In the dashboard, blackboard content appears in the work ledger and the "Context" tab of the agent drawer; messages appear in the chat stream and the "Messages" tab of the drawer.