Coding agents — Hermes, Pi, Omnigent

Plain chat answers. An agent acts — reads files, runs shell, plans, iterates — and streams every action back into the conversation.

Beyond plain LLM chat, Companion can hand a conversation to an agent. The principle is always the same: an external agent CLI runs on your machine (or a bridge host), Companion connects to it over a bridge, and the agent’s work streams inline below your messages. You stay in the chat; the agent does the doing.

Three agents, one mechanism — Hermes, Pi, and orchestrators like Omnigent. You summon one with a slash command; the conversation is then bound to it until you /exit.

Command	Agent	What it is	Configure
`/hermes [prompt]`	Hermes	Real-machine coding agent — shell, filesystem, browse.	Settings → Add-ons → Hermes Agent
`/pi`	Pi	Reflective conversational TUI — slow thinking, planning, journaling.	Settings → Add-ons → Pi
(bridge)	Omnigent	Multi-agent orchestrator (plan → act → verify), runs a whole software-factory loop.	bridged like the others; uses CoeOS as its model

/hermes — the real-machine agent

/hermes refactor auth.ts to use JWT instead of sessions
/hermes        ← enters the mode without sending anything yet

Hermes is a coding agent that runs on your workstation: it reads files, writes files, runs shell commands, browses — and streams its actions into an inline terminal panel under the messages. The mode is persistent: every message routes to Hermes until you /exit.

The panel never hides what the agent does:

Line	Meaning
`$ <your prompt>` (cyan)	what you sent
white text	agent tokens, streamed
`⚒ tool · <action>` (amber)	every file / shell action — nothing is silent
`⟲ reset`	drop the session; the next `/hermes` starts clean

The transcript is persisted — reload, switch conversations, come back, and the panel reappears with full history.

/pi — the reflective TUI

/pi

Pi opens as a terminal (TUI) embedded in the chat, for slow thinking, long-form planning, journaling. Unlike Hermes it isn’t a shell agent. Once in mode you type directly in the Pi terminal, not the main composer; /exit returns to normal chat. The Pi session persists across page reloads and survives switching conversations.

Omnigent — the orchestrator

Omnigent is a multi-agent orchestrator: it decomposes a task and drives a plan → act → verify loop across sub-agents (a software-factory workflow). It plugs into Companion the same way — bridged in, streaming inline — and it uses CoeOS as its plan-act model, so each phase is routed to the model your benchmarks proved best at it. Same principle as Hermes and Pi, one level up: an orchestrator instead of a single agent.

Rules that apply to all of them

One agent per conversation. You can’t run Hermes and Pi in the same chat at once.
Mode is per-conversation, not global. A second tab has its own mode.
Memory still applies. The agent sees what Némo knows — the memory snapshot is in the conversation’s system prompt.
Tools are always visible. Every file write, shell command and browse surfaces in the panel. No silent tool use.
/exit leaves any mode and returns to plain LLM chat; the agent transcript stays visible to scroll back.

Two directions

A coding agent inside Companion is one direction. The mirror is Agents tokens: mint an hms_… token and an external agent (Cline, Claude Desktop, Cowork) calls back into Companion’s memory and tools. Hermes can be wired both ways — driven from chat, and reading Companion’s brain over MCP.