Agents

See An Agent At Work

1. Open Agents.
2. Select an agent.
3. Use Chat to start work or continue a session.
4. Open the activity drawer to inspect the apps and tool calls used during the run.

What Are Agents?

The overview brings scheduled work, session totals, and current agent activity together. Use it to inspect the workspace before opening an individual session.

Agents are isolated AI assistants that reason through problems, execute tasks, and communicate across multiple platforms. Each agent has its own workspace, session history, model configuration, and channel bindings.

Agent Isolation

Every agent gets:

• Workspace, A dedicated directory for files, notes, and agent-specific data
• Sessions, Independent conversation history and transcripts
• Configuration, Model selection, tool permissions, spend limits
• Identity, A unique persona defined in the agent's AGENTS.md file
• Auth Profiles, Per-agent API key and authentication configurations

Agent Personas

Each agent's workspace contains an AGENTS.md file that defines personality, instructions, and behavioral guidelines. This persistent system prompt shapes how the agent responds, tone, domain expertise, formatting, tool usage, and safety policies.

Multi-Agent Routing

When running multiple agents, inbound messages are routed automatically based on bindings:

• Channel binding, Route all messages from a platform (e.g., all Telegram → personal agent)
• Server/workspace binding, Route a specific Discord server or Slack workspace to a dedicated agent
• Chat type binding, Route private messages to one agent and group messages to another
• Contact/peer binding, Route messages from specific contacts to specific agents

Bindings are evaluated in priority order. First match wins. Unmatched messages go to the default agent.

Sub-Agents & Delegation

Agents can spawn sub-agents for specialized tasks:

• Task delegation, Break down complex tasks and delegate parts to specialized sub-agents
• Isolated execution, Sub-agents run in their own session context
• Ping-pong conversations, Agents can have back-and-forth exchanges (up to 5 turns) to collaborate
• Result announcement, Sub-agent results can be posted back to the original channel

Teams & Orchestration

Multi-agent teams enable coordinated workflows:

• Parallel execution, Run multiple agents simultaneously on different aspects of a task
• Sequential pipelines, Chain agents where each builds on the previous output
• Hybrid mode, Combine parallel and sequential execution
• Task tracking, Monitor status of each agent's work (queued, running, done, error, canceled)

Tool Profiles

Control what each agent can do:

Custom allow/deny lists can be set per agent and per model provider.

Safe Mode

Safe Mode provides an execution sandbox:

• Master toggle, Enable or disable globally
• Per-agent controls, Disable for trusted agents, keep on for others
• Execution approval, Sensitive operations require explicit user approval
• Policy sync, Settings synchronized across all connected clients in real time

Agent Memory

Each agent has access to a memory system:

• MEMORY.md, Persistent notes and important information across sessions
• Memory files, Additional structured knowledge in memory/
• Vector search, Search memory using natural language queries

See Sessions & Memory for details.