Models and agent modes

Pilot supports multiple AI models and two agent modes that control how Pilot processes your instructions. This page describes the available options and when to use each one.

Available models

Pilot supports models from multiple providers. You can select a model from the More menu in the chat panel. Available model families include:

  • Claude (Anthropic): Claude Opus and Claude Sonnet variants
  • GPT (OpenAI): GPT model variants
  • Gemini (Google): Gemini Pro variants

Different models have different strengths. In general, larger models produce higher-quality output for complex tasks, while smaller models respond faster for straightforward changes. You can switch models at any time during a session without losing your work.

The model selector in the More menu showing available model options.

Agent modes

Pilot operates in one of two modes, controlled by the mode toggle next to the prompt input.

The agent mode selector in the chat panel.

Act mode

Act mode is the default. When you submit a prompt, Pilot executes changes immediately. The agents read your instructions, make decisions, and apply changes to the ontology, design, and code without waiting for intermediate approval.

Use act mode when you want Pilot to work autonomously and you are comfortable reviewing changes after they are applied.

Plan mode

In plan mode, Pilot creates a detailed plan before executing changes. After you submit a prompt, Pilot outlines what it intends to do and waits for your approval before proceeding.

Use plan mode in the following situations:

  • You want to review proposed changes before they are applied.
  • You are making significant structural changes to the ontology or application.
  • You want to understand the scope of a change before Pilot begins work.

To toggle between modes, select the mode toggle next to the prompt input in the chat panel. The current mode is displayed as Act or Plan.

Subagents

Pilot delegates work to specialized subagents, each focused on a specific aspect of the application. The subagents are as follows:

  • Ontology architect: Defines and modifies object types, action types, link types, and their properties.
  • Designer: Creates the design specification, including color palette, typography, layout, and interaction patterns.
  • App builder (frontend architect): Implements the React application using the ontology and design specification.
  • Seed data generator: Creates realistic sample data for testing the application in a container.

When a subagent is active, the chat panel displays its activity in a separate thread. You can expand the thread to follow the subagent's reasoning and tool usage.

Agent tools

Pilot agents use a set of tools to build and modify your application. You do not need to invoke these tools directly. Pilot selects and uses the appropriate tools based on your instructions.

Pilot uses agent tools automatically. Interact with Pilot through natural language prompts, and Pilot determines which tools to use.

The following tools are available to agents:

  • File operations: Reading, writing, and editing source code files
  • Search tools: Finding files and searching file contents by pattern
  • Shell commands: Running build commands, installing dependencies, and executing scripts
  • Ontology search: Querying existing ontology object types, action types, and functions
  • Version control: Committing and pushing code changes