Agents, tools and skills
An agent is a model given instructions, knowledge and permission to act. In Olympusx AI an agent can answer from collections, call tools on servers you register, follow skills that tell it how to do a task, and hand questions to collaborator agents. Every step it takes is authorized, bounded and recorded. This page explains the parts and how they combine.
The parts of an agent
Section titled “The parts of an agent”| Part | What it is |
|---|---|
| Model | The granted, deployed model it thinks with. An agent with tools, skills or collaborators needs a model that can call tools. |
| Instructions | What it is for and how it should behave. |
| Routing key | A short description of when to use this agent, at most 280 characters. Other agents read it to decide whether to hand it a question, so it says when, not how. |
| Knowledge | The collections it answers from. See Knowledge and grounding. |
| Tools | Tools on servers the tenant registered, that this agent is granted. |
| Skills | Named procedures it can load when a task calls for one. |
| Collaborators | Other agents in the tenant it may hand a question to. |
| Ceilings and approvals | How much one run may do, and which tools wait for a person. |
A tool is an action an agent can take outside itself: look up a record, open a ticket, query a system. Tools live on tool servers that speak the Model Context Protocol over streamable HTTP. A tenant registers a server once; each agent is then granted the tools it may call, one by one. An agent cannot call a tool it was not granted, whatever the model asks for.
Olympusx AI reaches a tool server only at an address it is allowed to: public addresses over https://, and private addresses only when the platform administrator allow-lists them. Cloud metadata and loopback addresses are never reached. A server’s credential is sealed when it is registered and never shown again. See Tools.
Skills
Section titled “Skills”A skill is a procedure an agent loads when a task needs it: instructions for one kind of job, and optionally the tools that job needs. While a skill is in use, the agent may call only the tools that are both granted to the agent and named by the skill. A skill narrows what an agent can do; it never adds a tool the agent was not granted. See Skills.
Collaborators
Section titled “Collaborators”A collaborator is another agent the agent may hand a question to. Olympusx AI offers each collaborator to the agent as a tool, described by its routing key, and the model decides when a question belongs to one. The collaborator answers with its own model, instructions, knowledge, tools and skills.
Delegation is bounded: an agent has at most 12 collaborators, and a question is handed on at most twice from the agent the person spoke to. See Collaborators.
Ceilings
Section titled “Ceilings”Every run has ceilings, so an agent that loops or wanders stops instead of spending without end.
| Ceiling | Default |
|---|---|
| Tool calls in one run | 20 |
| Model calls in one run | 25 |
| Tokens in one run | 200,000 |
| Wall-clock time for one run | 5 minutes |
| One tool call | 60 seconds |
| One tool result the model reads | 32 KB; a longer one is cut and the model is told |
A builder can set the first three per agent, and a run started over the API can set them per run. A run that reaches a ceiling stops and records which one. The other limits are set by the platform administrator. See Limits.
Approvals
Section titled “Approvals”Any tool can be marked to wait for a person. When the model asks for that tool, the call does not run: the run stops, and an Admin or Reviewer decides. See Human approval.
How agents reach people
Section titled “How agents reach people”An agent is built and tried in Build › Agents, then published as an assistant for people to use: the tenant’s own people in the console, or anyone with the assistant’s share link. An assistant answers through its agent’s tools, skills and collaborators, and holds a conversation until a pending approval is answered. See Assistants and Agents and assistants over the API.
What is recorded
Section titled “What is recorded”Every run is kept with each step: the model calls, the tools called with their inputs and outputs, the hand-offs to collaborators, the approvals, the tokens and the outcome. Runs are in Operate › Runs, and their traces in Operate › Traces. See Observability and evaluation.