What is an agent?
You give an agent instructions, a model, and access to selected tools. It can answer questions, search the web, work on a schedule, send messages, and store information when those features are configured. Each agent has a logical identity and normally uses agent-scoped session keys, memory and context queries, workspace paths, instructions, budgets, and tool policy. You can have one agent handling your Discord server and another managing your Telegram chats, each behaving differently based on its configuration. This scoping is not process isolation. Agents share the same trusted daemon and host, and explicitly authorized cross-agent operations can cross those logical boundaries. Use separate deployments when agents belong to mutually hostile security domains.What agents can do
Identity
Give your agent a name, personality, and set of instructions using
workspace files — simple Markdown documents that shape who your agent is.
Lifecycle
Understand the journey of a message from arrival to response, including
memory recall, tool use, and safety checks along the way.
Memory
Store and recall selected information with source, trust, and correction
history when memory and learning are enabled.
Tools
Agents can search the web, manage schedules, send cross-platform messages,
analyze images, and much more through built-in and custom skills.
Safety
Use token limits, circuit breakers, step counters, graph budgets, and opt-in
spend ceilings. Configure and test the controls your workflow needs.
Models
Catalog-driven access to the current providers in the pi-ai SDK (see
comis models list
for the current count — Anthropic, OpenAI, Google, Mistral, Groq, xAI, Cerebras,
OpenRouter, DeepSeek, Bedrock, Vercel AI Gateway, and more), plus custom OpenAI-
compatible endpoints. Switch models per agent, set up automatic
failover, and rotate API keys.How it all fits together
When someone sends a message to your agent, here is what happens at a high level:- The message arrives from any connected platform (Discord, Telegram, Slack, WhatsApp, or the web dashboard)
- Comis routes it to the right agent based on your routing rules
- The agent recalls relevant memories from past conversations
- The agent thinks and optionally uses tools (web search, scheduling, file management, and more)
- A reply is sent back to the same chat
Slash commands
While chatting, you can control your agent with slash commands — short instructions that start with/. For example:
/think high— Tell your agent to think harder about the next response/status— See session info, token usage, and model details/model openai/gpt-4o— Switch to a different AI model for the session/new— Start a fresh conversation
/skill:name dispatch is not wired in
the current daemon. See Slash Commands.
Running multiple agents
You do not need routing rules if you only run one agent — the default agent handles everything automatically. But if you want different agents for different situations, Comis supports that out of the box. For example, you might run:- A support agent on Slack that answers team questions
- A research agent on Telegram that finds and summarizes information
- A scheduler agent on Discord that manages reminders and recurring tasks
Explore agents
Agent Lifecycle
How your agent processes messages from arrival to response.
Identity
The workspace files that shape your agent’s personality.
Models
Choosing and configuring AI providers for your agents.
Safety
Budget limits, safety fuses, and cost protection.