Skip to content

Agents: Core Platform

Evonic is fundamentally an agentic AI platform. Agents are the core building blocks that power everything — from simple chatbots to complex multi-agent workflows. With Evonic, you design agents from concept to production, defining every aspect of their behavior: the underlying model, available tools, knowledge base, communication channels, and installed skills.


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

  • Reason and plan — Break down complex tasks into actionable steps
  • Take actions — Execute tools, run code, query databases, and call APIs
  • Remember context — Maintain conversation history across sessions with persistent state management
  • Communicate — Talk to other agents natively via the agent-to-agent protocol
  • Connect anywhere — Deploy to Telegram, WhatsApp, Discord, or web interfaces
  • Learn from knowledge — Access reference documents and custom knowledge bases
  • Operate anywhere — Run on local machines, remote SSH servers, or tunnel devices via Workplaces

When you design an agent in Evonic, you define these components:

ComponentDescription
ModelThe LLM that powers the agent’s reasoning — OpenAI-compatible, local, or cloud
System PromptThe agent’s personality, role, and behavioral guidelines
ToolsExecutable capabilities (bash, runpy, web search, custom API tools)
Knowledge BaseReference files the agent can read during conversations
ArtifactsPersistent file output system for saving work results across sessions
ChannelsMessaging platforms the agent connects through (Telegram, WhatsApp, etc.)
SkillsInstallable packages that bundle tool definitions with Python backends
WorkplaceExecution environment — local directory, SSH server, or Evonet-connected device

Evonic’s architecture is defined by three differentiating concepts — Workplaces, Agent-to-Agent Communication, and HMADS (Heuristic Mal-activity Detection System) — which are explained in detail on the Design Philosophy page.


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

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

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

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.

Each agent has its own artifacts directory (agents/<agent_id>/artifacts/) for persisting work outputs across sessions. Agents use the save_artifact tool to store reports, images, or any generated files. Artifacts survive session deletion and are accessible through the web UI. See Artifacts for details.

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.

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 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. See Channels for details.

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.

Introduced in v0.7.0.

Press Ctrl+G anywhere in the web UI to open a keyboard-driven overlay for instant agent search and navigation. Type a partial name to filter agents, then jump directly to any agent without touching the mouse or leaving the current page.

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

CommandDescription
/clearClear the current conversation history
/clear-memoryClear the agent’s long-term memories
/helpDisplay available slash commands
/summaryTrigger conversation summarization
/execSwitch from plan mode to execute mode
/shutdownShut down the Evonic server (super agent only)
/cwdShow the current workspace directory
/cdChange the agent’s workspace directory

See Slash Commands for full details.


Evonic agents excel in a wide range of scenarios. Here are some of the most common:

CategoryUse CaseDescription
Customer ServiceSupport AgentHandle tickets, FAQs, refunds, and escalations
PersonalPersonal CompanionDaily tasks, reminders, research assistance
Multi-AgentAgentic SwarmCollaborative agents working together
DevOpsInfrastructure AgentServer monitoring, deployment automation
DataResearch & Analyst AgentData extraction, summarization, reporting
HRRecruitment AgentCandidate screening, interview scheduling
EducationTutor AgentPersonalized learning, Q&A, assessments
HealthcareTriage AgentSymptom assessment and appointment scheduling
E-CommerceShopping AssistantRecommendations, order tracking, returns
FinanceFinancial AdvisoryPortfolio analysis, market summaries
EnterpriseAgentic ERPSupply chain, inventory, procurement, finance
ML/AIAI Workflow OrchestrationEnd-to-end ML pipeline automation

For detailed walkthroughs of each use case, see the Use Cases page.


  1. Navigate to /agents in the web UI
  2. Click + New Agent and enter an ID (e.g., my_assistant)
  3. Configure the system prompt in the General tab
  4. Select the model for the agent’s reasoning
  5. Upload knowledge files in the Knowledge tab
  6. Assign tools and skills in the Tools tab
  7. Test immediately using the Chat tab
  8. Optionally connect a Telegram or WhatsApp bot in the Channels tab