Agents, Bots and Subagent roles

The Main Agent is rho's persistent identity for Web conversations. A Bot is a separate persistent identity with its own name, purpose, welcome message and selected Agent Preset. The Preset owns instructions, model selection, fallback models and tools. Sharing a Preset does not share conversation history or private memory.

Nexus holds each Bot as a named Agent under your account, with an @<name> handle and no separate credential or executor address. The Bots page owns each Bot's current conversation, history, Home Workspace and channel settings. A Bot can use a shared Telegram connection with one fixed private-chat, group or topic destination.

The Workspace owner sets each Bot's messaging mode to Public or Private. Public accepts contact from every otherwise eligible Agent; Private accepts only its allowed senders and itself. Each Bot has its own setting, including Bots served by the same rho instance.

Create or edit a Bot

1. Open Bots in the sidebar and select New bot, or open an existing Bot's Settings.
2. Enter its name, display name and description. The name becomes its handle and cannot be changed after creation.
3. Choose Personal or External requests as its purpose.
4. Select an Agent Preset. A new Bot starts with the New Bots default selected in Settings › Agent Presets.
5. Leave Subagent on to allow background workers, or turn it off to make the Bot complete work directly in its main conversation. It starts on for a new Bot.
6. Choose a default, custom or disabled Introduction, then select Create bot or Save.
7. Save a Home Workspace before creating its first conversation. Configure its channel target on the same page.

Edit selected Preset opens that Preset's configuration. Changes there apply to new work for every identity using it. Already accepted work keeps its captured configuration. The Bot form contains no independent prompt, model or tool settings. Its model and tool count on the list are read-only projections of its effective configuration.

Bots always keep Code mode off. A Bot's own Subagent setting applies to future inputs in its current, new, resumed and forked conversations. Main's global defaults and conversation buttons do not change it. With Subagent off the Bot waits for the work it performs directly and then reports the result. Existing workers can still finish and report; their captured tools, including Code mode, remain unchanged.

The introduction is static text stored with the Bot. The default uses its name and description; custom text replaces it, and None disables it. Editing it does not send a message or make a model request.

Choose Main's Preset in Settings › Agent Presets › Default Presets; Settings › Defaults owns rho's global initial model and ordered fallback list. The main Agent has no separate built-in prompt replacement editor.

A Bot with no effective model is marked No model and cannot answer requests. Edit its Preset or the global model setting to supply one. Model recovery uses the execution's captured ordered candidates and constraints; the global initial model is not an implicit fallback. A plugin-owned Agent is marked Plugin and uses the plugin's workflow.

Duplicate or remove a Bot

Duplicate creates a fresh identity with the same Preset reference and welcome message and Subagent setting. Choose a new name and purpose. It copies no memory, conversations, Home Workspace, Telegram credentials or chat bindings.

Remove bot stops its running work after confirmation. Its conversations stay readable but cannot continue, and its name remains reserved. The selected Preset remains available. The main Agent cannot be removed here. Import and export belong to Presets, so a configuration file never creates or overwrites a Bot identity.

Purpose and memory

Choose Personal for a Bot that serves you, or External requests for one that serves admitted visitors. This purpose is separate from the Workspace's Public/Private messaging setting: purpose limits what the Bot can use, while messaging controls which Agents may contact it.

An external Bot can write conversation/ and speaker/ memory and read workspace/ knowledge. It has no Agent-private memory, Human-owned memory, owner persona, Runner tools or MCP tools. Its conversation tools and history stay within its current conversation and descendants. It delegates through the published Subagent roles, with the same external-purpose limits applied to every child. Separate peer messages may address itself or an eligible external Bot from the same rho instance using the exact UUID in the send tool. Workspace access and messaging rules still apply.

A group Bot shares one current conversation and history among all admitted participants. Its owner controls new/resume and other shared lifecycle changes. All Bots use the same Speaker memory mechanism. speaker/ holds one Bot's notes about one person; another Bot has separate notes about that person. A private Bot's owner is simply its usual Speaker. Subagent work retains attribution but does not automatically read that Speaker's or the owner's long-term notes.

An external identity cannot become private later. Use Duplicate, give the copy a new name, and choose its purpose. The copy receives a new identity and retains its Preset reference, welcome message and Subagent setting; it has no copied memory, conversations, Home Workspace, Telegram credentials or chat bindings. A private Bot can become external; new work uses the restricted policy, while accepted work retains the policy it captured.

Commands

rho agents                     # list persistent Bots and their effective model
rho agents duplicate reviewer personal-reviewer owner
rho agents rm reviewer         # stop and remove this Bot; its name stays reserved
rho bots current reviewer      # select its current conversation, creating one if absent
rho bots new reviewer          # create and select a conversation in its Home Workspace
rho bots history reviewer      # conversation UUIDs and their recorded Workspace UUIDs
rho bots history reviewer --after CURSOR --limit 20
rho bots resume reviewer CONVERSATION_ID --workspace WORKSPACE_ID

A Bot selector is its name, @handle or public UUID. Set its Home Workspace in the Bot's settings before opening its first conversation. Resume takes the Workspace printed with that history row, even after the Bot's Home changes. New and Resume leave accepted work running in its original conversation. See Bot conversations for their archive behavior.

Each switching command prints an operation ID before sending the request. After an interrupted or uncertain result, retry the same command with --operation-id ID; reusing that ID does not perform the switch again. If another command has since selected a different conversation, the output shows both the original selection and the current one. History only reads.

A plugin-provided Agent is managed by its plugin. The Bot editor does not change it.

Publish and use Subagent roles

Open Settings › Subagent roles. The initial roles are worker, planner, explorer and reviewer. Each publication names an existing Preset and adds the role name and description used to select it. Edit Preset opens its shared configuration; the role editor has no second prompt, model or tool configuration. Unpublishing removes a role from future delegation catalogs without deleting its Preset or interrupting already accepted work.

The Main Agent and Bots select an exact role when delegating. Ordinary work uses worker; a later review is a separate reviewer task with the implementation result and relevant evidence in its brief. Main Agent and Bot identities cannot be used as spawn targets, including through a direct UUID or handle.

A Subagent starts a child Conversation with the task brief and the role's captured configuration. It uses the caller's identity without creating another User. Its memory is limited to its own Conversation and its conversation tools to that Conversation and descendants. It receives neither the parent's full transcript nor owner or Speaker notes automatically; include the context it needs in the brief. The selected role may use a different subset of the installation's tools from its parent, within the same Bot-purpose limits.

Follow the initiating conversation takes the calling task's current model, including any completed fallback. Use the Preset's initial model takes that Preset's resolved default. A model explicitly requested for one delegation overrides either choice for that task; it does not edit the Preset. The role's fallback list remains independent. Later unqualified messages to the child use the child's own current model and captured configuration. Those defaults also survive a fork or cleanup of the parent's execution details.

Read work in progress

The agent can inspect a readable conversation while its Bot or sub-agent is still working. status reports whether the conversation is busy; session_read reads its context. With include_execution: true, it also returns the current execution's state, persisted intermediate text and tool output previews, including tools that have not finished:

{"session_id":"<worker conversation UUID>","include_execution":true,"limit":1}

The worker UUID comes from spawn. The latest returned turn is inspected by default; around_turn_id selects a particular turn and keeps later execution pages attached to it. A later read observes newly persisted progress. This does not stop the worker or wait for it to finish. A Bot needs session_read in its selected tools, and its usual workspace and conversation restrictions still apply; external Bots remain within their conversation and descendants.

This read is separate from Side's fixed reference snapshot. It provides bounded previews of recorded progress, with pagination for older model rounds; live token streaming remains on the existing feeds. See the history reference for limits, regeneration candidates and retained execution details.

Where this comes from

A named Agent is a kernel resource (Named agent definitions). rho owns the product's Main Agent, Bot and Preset distinctions and compiles the selected Preset into Nexus execution declarations. Nexus retains identity, execution, memory and permission authority.