Sapphire can spawn background AI agents that work independently while you keep chatting. They run their own missions, use their own tools, and report back when done.

Think of agents as AI workers you dispatch on tasks. While you're talking to Sapphire about your day, an agent can be researching something, writing code, or analyzing data in the background.

When all agents from a chat finish, their results are delivered back to that chat automatically.

Just ask Sapphire to spawn one:

• "Spawn an agent to research quantum computing"
• "Send an agent to summarize my last 50 emails"
• "Dispatch two agents — one for code review, one for documentation"

The AI uses these tools behind the scenes:

The default. Runs an LLM with tools in a background thread — same capabilities as a normal chat but isolated.

• Has its own prompt, toolset, and model
• Can use tools (memory, web search, etc.)
• Up to 10 tool rounds per mission
• Scopes set to 'none' by default (clean isolation)

Spawns a Claude Code session for coding tasks. Requires claude CLI installed on your system.

• "Spawn a Claude Code agent to fix the login bug in my project"
• Works in a configurable workspace directory
• Three execution modes: strict (file ops only), standard (code + run), system_killer (unrestricted)
• Sessions can be resumed

Active agents show as colored pills above the chat input:

• Blue — pending/starting
• Yellow — running
• Green — completed
• Red — failed or cancelled

Click a pill to see the agent's status and result.

1. You spawn one or more agents from a chat
2. They work in the background
3. When an agent finishes, it waits for the whole batch
4. Once all agents from that chat are done, results are delivered as a message in that chat
5. If you're mid-conversation when results arrive, they queue until there's a break

This means you get a consolidated report, not a stream of interruptions.

Configure in Settings → Plugins → Agents:

Agents are limited by max_concurrent. If you try to spawn more than the limit, you'll be told to wait. Agents from different chats all share the same pool.

• "Spawn an agent to research the best Python web frameworks in 2026"
• "Send an agent to check all my SSH servers for disk usage"
• "How are my agents doing?"
• "What did the research agent find?"
• "Cancel the documentation agent"
• "Dispatch a Claude Code agent to write unit tests for the auth module"

• Agent stuck on pending — Max concurrent limit reached. Wait for others to finish or dismiss one
• Agent failed — Check the error in the agent's report (recall it). Common: bad toolset, model unavailable
• Results not showing — They deliver to the chat the agent was spawned from. Check that chat
• Claude Code agent failed — Make sure claude CLI is installed and accessible

Background agent system for parallel AI task execution.

TOOLS:

• agent_options() - list available types, models, toolsets, prompts
• spawn_agent(mission, agent_type?, model?, toolset?, prompt?, project_name?, session_id?) - launch agent
• check_agents() - status of all active agents
• recall_agent(agent_id) - get completed agent's report
• dismiss_agent(agent_id) - cancel/cleanup agent

AGENT TYPES:

• llm: background LLM + tool loop (default). spawn_args: model, toolset, prompt
• claude_code: Claude Code CLI session. spawn_args: project_name, session_id

LIFECYCLE:

• pending → running → done/failed/cancelled
• Batch completion: all agents from same chat finish → consolidated report delivered
• Auto-return to originating chat, queues if mid-stream

LIMITS:

• max_concurrent (default 3, configurable)
• LLM agents: max 10 tool rounds per mission
• Scopes default to 'none' for isolation

EVENTS (SSE):

• agent_spawned, agent_completed, agent_dismissed, agent_batch_complete

UI:

• Pill bar (#agent-bar) above chat input
• Color-coded: blue=pending, yellow=running, green=done, red=failed