Skip to content

Agents, models, and skillsets

Three independent selections shape a Nori session: the agent that performs the work, the model it uses, and the skillset that supplies team-specific instructions and tools.

The agent is the coding product launched inside the session machine. The current first-class Sessions set is:

  • Claude Code
  • Codex
  • Cursor Agent
  • OpenCode
  • pi-acp

Some agents are already installed in the base environment. Others are installed when first launched. That delivery detail does not change the user-facing support status.

Gemini CLI is supported by the open-source Nori CLI but has been removed from managed Sessions. This distinction prevents a local CLI capability from being mistaken for a hosted product guarantee.

The model is resolved within the selected agent’s supported catalog. A fleet can set a default model per agent, and interactive surfaces can request a different valid model.

Model availability can also depend on the connected account or plan. Nori’s catalog validates selections, but it cannot make a model available to an account that the upstream provider does not permit.

A skillset, also called a personality in parts of the product, is the package of instructions and configuration applied to the agent when a session is prepared. It can include:

  • Agent instructions and operating rules.
  • Reusable skills and their supporting files.
  • MCP configuration.
  • Agent-specific files generated from one common source.

The fleet can define a default skillset. A session may select another when it starts, Slack can switch skillsets with !skillsets switch, and a trigger can pin a skillset for unattended runs.

At session start, the broker resolves the requested values against organization policy:

  1. Determine the agent from a request override or the fleet default.
  2. Confirm that the agent is enabled and supported.
  3. Resolve a valid model for that agent.
  4. Resolve the requested or default skillset.
  5. Prepare the machine and launch the agent with those settings.

Changing the agent mid-conversation can move the logical conversation to a fresh machine. Transcript context carries across, but uncommitted working files do not automatically follow an agent switch.

The compute backend determines where the machine runs. The agent and model determine what runs inside it. A fleet’s compute choice does not restrict the conversation to one first-class agent.

For credentials and exact availability, see supported agents and compute and integrations and credentials.

this page as markdown: /sessions/concepts/harnesses-and-personalities.md· index for agents: /llms.txt