Agents — Core Platform

The Evonic AI is fundamentally an agentic AI platform. Agents are the core building blocks that power everything — from simple chatbots to complex multi-step workflows.

What is an Agent?

An agent is an independently configured LLM-powered assistant that can:

Reason and plan — Break down complex tasks into steps
Take actions — Execute tools, run code, query databases
Remember context — Maintain conversation history across sessions
Connect anywhere — Deploy to Telegram, WhatsApp, Discord, or web interfaces
Learn from knowledge — Access reference documents and custom knowledge bases

Architecture

User Message
    ↓
Channel (Telegram, Web, etc.)
    ↓
Agent Runtime
    ├─ Load agent config (system prompt, model, tools)
    ├─ Load/create session (per-user persistence)
    ├─ Build messages (system prompt + history + new message)
    ├─ Call LLM
    ├─ Execute tool calls (if any)
    ├─ Loop until final response
    └─ Save messages to session
    ↓
Response → Channel → User

Key Concepts

Agents

Each agent has a workspace directory (agents/<agent_id>/workspace/) for file operations. This is configurable per-agent in the General tab and provides an isolated filesystem context for tools like runpy, bash, and write_file.

For agents that need to operate on remote machines or devices, assign a Workplace — a first-class execution environment that can be a local directory, an SSH server, or any device running Evonet (a lightweight Go connector). See Workplaces for details.

An agent is defined by a slug ID (e.g., bookstore_bot, library_assistant), a system prompt, an optional model override, and a set of assigned tools. Manage agents at /agents in the web UI.

Knowledge Base

Each agent has a filesystem directory (agents/<agent_id>/kb/) for storing .md reference files. The agent can read these files using the built-in read tool during conversations. This keeps knowledge files editable and version-controllable.

Tools

Agents can use tools from the tool registry. In production mode, tools call real Python backend implementations (in backend/tools/). The built-in read tool is always available and scoped to the agent’s KB directory.

Agent State

Agents operate in one of two modes: Plan or Execute. The agent state is persisted across conversation turns and survives LLM context summarization. Agents always start in Plan mode on a new session. See Agent State for details.

Channels

Channels connect agents to external messaging platforms. Each channel type (Telegram, WhatsApp, etc.) implements a common interface. Multiple channels can be connected to the same agent.

Sessions

Conversations are persisted per-user, per-agent, per-channel. The same user can have separate conversations with different agents. Session history is used as context for each LLM call.

Slash Commands

Agents support slash commands for quick actions. These work in both the web chat and Telegram channel:

Command	Description
`/clear`	Clear the current conversation history
`/help`	Display available slash commands
`/summary`	Trigger conversation summarization

See Slash Commands for full details.

Quick Start

Navigate to /agents in the web UI
Click + New Agent and enter an ID (e.g., my_assistant)
Configure the system prompt in the General tab
Upload knowledge files in the Knowledge tab
Test immediately using the Chat tab
Optionally connect a Telegram bot in the Channels tab