Glossary

Agent Runtime

The software that actually runs an agent: it loops the model, executes its tool calls, keeps its memory, decides when it wakes up, stops it when it runs too long and records what it did.

General definition

An Agent Runtime is to an AI agent what a server is to a web application: the environment that executes it. An AI agent is defined by a model, a system prompt, a set of tools and some memory, but none of that does anything until a runtime sends the prompt to the model, reads the response, runs the tools the model asked for, feeds the results back, repeats until the task is done, and stores what happened. Frameworks help you write the agent; the runtime is where it lives in production.

  • The loop: prompt, model response, tool calls, results, next turn, with a stop condition
  • Memory and state: conversation history, long-term memory, intermediate results, and persistence across restarts
  • Scheduling and triggers: agents that start on a message, a webhook, a timer or idle time, not only on a user prompt
  • Limits: maximum turns, token and cost budgets, timeouts, and gates on which agent may respond to which event
  • Observability: traces of every model call and tool call, errors, latency and spend, exported to monitoring
  • Isolation: permissions on tools, credentials and network, often with a sandbox for code execution

Runtimes come in three shapes. Hosted agent platforms from model vendors and cloud providers run the loop as a managed service and bill per call or per hour. Workflow and durable-execution engines run agents as long-lived jobs with retries and checkpoints. Self-hosted runtimes run inside your own application or server, which is the only option when the agent must act on data that may not leave your network. The agent orchestration layer, which coordinates several agents, sits on top of whichever runtime is chosen; a multi-agent system needs a runtime that can run several loops and arbitrate between them.

The practical questions when choosing one: where does state live and can you inspect it, how are runaway loops stopped, how are tool credentials scoped, can an agent be triggered without a human message, and what do the traces look like at three in the morning when the agent did something wrong.

In the Ethora ecosystem

In Ethora the agent runtime is the chat server itself. Agents built with the AI SDK are members of rooms and channels, so the trigger for an agent turn is a message in a room it belongs to, and its memory is the room: a per-room transcript log serves as working memory and the knowledge base serves as long-term memory through retrieval. Tool calling is built in, and the MCP server exposes apps, users and rooms as tools and resources for Claude and other MCP clients.

The limits are explicit. In multi-agent rooms a response gate decides which agent answers a given message, and an autonomous-turn cap stops agents from replying to each other indefinitely. A heartbeat scheduler lets an agent speak first on a schedule or after a period of idle time, and a one-shot reply endpoint runs a single agent turn from your own code without a room event. Because the runtime is the server, everything an agent says lands in the server-side archive and the audit trail, the Trust and Safety system screens its messages like anyone else’s, and on a self-hosted or dedicated deployment the loop, the memory and, with a local model, the inference all run inside your network.

Get started

Build AI agents on your own stack

Ethora’s AI SDK brings agents, RAG and self-hosted LLMs into your product. Talk to our team.

Start Free
Free tier available Enterprise SLA No vendor lock-in