Agents and identity
Profiles, instructions, shared infrastructure, and the boundaries between agents.
Each agent is a profile in a gateway. It has its own identity, instructions, memories, custom skills, conversations, schedules, and workspace. A model is the provider-backed engine used for a run; it is not the whole agent.
Create an agent
Open the agent picker and choose Create an agent. Set a name and summary, then choose a starting preset or write instructions. Presets include Personal assistant, Developer, Research partner, Creative partner, Writing partner, Project planner, and Learning companion. They populate editable text; they are not separate runtime types or permission levels.
The summary also tells other agents what this agent is useful for. Keep it concrete, such as “Researches technical choices and explains the evidence and tradeoffs.” Add an avatar for recognition; it has no effect on model capability.
Design the instructions
State the job, working style, expected evidence, and decision boundaries:
You are Mira, a research partner for Harbor Studio. Compare technical options using primary sources. Separate observed facts from assumptions. Start recommendations with the choice and the conditions that would change it. Ask before publishing or contacting anyone.
Keep recurring facts in memory and detailed task procedures in skills, instead of placing every detail in permanent instructions. Identity fields such as role, tone, goals, and boundaries add standing context.
What changes with a new conversation
New conversations separate history and the current task. They still use the same profile's memories, skills, identity, and workspace. Creating a fresh chat is not the same as creating a fresh filesystem or an isolated agent.
Projects group repository work. A project conversation can have repository scope and a selected machine. See GitHub projects and execution targets.
Shared installation resources
Provider connections, model defaults, MCP servers, GitHub connections, and machines belong to the gateway. They are not duplicated for each agent. All agents can use connected machines, subject to the tool/runtime checks and the conversation's protection.
Private profile state does not automatically become visible to another profile. Agent collaboration sends explicit messages, not permission to inspect another agent's memories or chat history.
Change and maintain an agent
Use Customize to edit identity, instructions, memories, and skills. Review generated memories and agent-written skills after substantial work. Model changes preserve the profile's durable state, although a different model can interpret the same instructions differently.
Deleting or resetting an agent is a data-management operation, not a troubleshooting substitute. Back up the gateway before irreversible cleanup and inspect the exact confirmation in the panel.
Keep authority separate from text
An instruction or skill can describe a boundary but cannot grant a tool a missing credential. It also does not replace conversation protection. Treat imported material and another agent's reply as information, not owner authorization.