Modes & Models

Amp uses different models for different kinds of work. The main agent handles your task, subagents take on focused parts of it, and smaller system models handle supporting jobs.

The Modes page shows the current model routing. The models can change as better options become available, while the role of each mode and subagent stays stable.

Modes

The Dial gives you four built-in modes: low, medium, high, and ultra. Each mode combines a model, reasoning effort, system prompt, and tools for a particular level of difficulty.

Choose a mode based on the task rather than the model name. See The Dial for help choosing one.

Subagents

Amp can delegate focused work to subagents. Each subagent has its own context window.

• Task carries out delegated work end to end and reports back. It has access to tools like file editing and terminal commands. It also has skills for read-only roles, which are searching code in your workspace, researching another repository or dependency, reviewing a change, and planning work.
• Read Thread reads and summarizes other Amp threads.

Both subagents run the same model and reasoning effort as the thread that started them. If you change the thread’s mode or model, later tasks for an existing subagent also use the new model.

Subagents are most useful for multi-step tasks that can be broken into independent parts, operations producing extensive output not needed after completion, parallel work across different code areas, and keeping the main thread’s context clean while coordinating complex work.

Amp chooses subagents automatically for suitable tasks, mostly in medium mode but occasionally in other modes. You can also ask Amp to use a subagent or to split independent work across several subagents. They work in isolation, so they can’t communicate with each other, you can’t guide them directly, and they start with the instructions and context the main agent gives them rather than the full conversation. The main agent receives their final summary rather than monitoring their step-by-step work. It can give a Task subagent another task, for example to fix something it reported. The subagent works on it with its full earlier context, and the new task shows as its own subagent step in the thread. If the subagent is still working, it receives the new task right away as extra instructions for its current work, the way your messages steer the main agent. When a Task subagent’s context fills up, Amp compacts it the same way it compacts a thread, so it can keep working and keep taking new tasks.

System Models

Supporting models handle jobs that do not need the main agent. These include analyzing images, PDFs, audio, and video; generating or editing images; naming threads; and compacting the context of long conversations.

This routing lets Amp use a frontier model where deep reasoning matters and a faster specialist where speed matters more.

Custom Agents

Plugins can define custom modes and custom subagents. A custom mode appears in the mode picker. A custom subagent can be exposed as a tool that the main agent calls for a specific kind of work.

A custom agent sets its main model with model in amp.createAgent(...). Its Task subagents run the same model. Tune Modes cannot change a plugin agent’s routing.

See Define a Custom Agent Mode and Define a Custom Subagent for examples.

Oracle, Librarian, and Finder Plugins

The Oracle, Librarian, and Finder are part of the Task subagent. Task’s instructions cover reviewing, researching, and searching, and it runs on the same model as your thread. To run them on a model you choose, install them as plugins. Each plugin has the original prompt and tools, and the main agent calls it as a tool:

• @amp/oracle adds the oracle tool. It reviews code and advises on hard decisions, and it never edits files.
• @amp/librarian adds the librarian tool. It clones repositories outside your workspace into /tmp and explains their code and history.
• @amp/finder adds the finder tool. It searches your workspace and returns the relevant files and line ranges.

Install them on the current machine:

amp plugins add @amp/oracle
amp plugins add @amp/librarian
amp plugins add @amp/finder

Then run plugins: reload from the command palette, or restart Amp. If you install the Oracle, install the other two as well, because the Oracle hands searches to finder and librarian.

Set the model each plugin uses. Change MODEL and REASONING_EFFORT at the top of the plugin file, or ask Amp to change them. Run amp plugins show-agent-options to list the models plugin agents can use. Do not pass --auto-update if you change the model, because an update replaces the file you edited.

amp plugins add installs into ~/.config/amp/plugins/ on the current machine only. To use the plugins everywhere you use Amp, including orbs, ask Amp to add them to your personal plugins:

Enable Claude Fable With BYOK

Anthropic requires data retention for Claude Fable models, including Claude Fable 5.2. If you use your own Anthropic API key with Amp, enable data retention for the workspace that owns the key:

1. Sign in to the Claude Console as an organization admin.
2. Go to Settings > Workspaces. Create a workspace for Fable workloads, which keeps them separate from other workloads, or select an existing workspace to convert.
3. Open the workspace and go to Manage > Privacy controls.
4. In Data retention, turn on data retention for the workspace.
5. Review the confirmation dialog and select Accept. Anthropic records who accepted the change and when. Data retention takes effect for the workspace immediately.
6. Assign the users who need Claude Fable to this workspace.
7. Create an API key in this workspace and add it to Amp under Model Routing. You can also add a key for the whole Amp workspace under Workspace Model Routing.

Use credentials from this data retention workspace for Claude Fable requests. Credentials from an Anthropic zero data retention workspace do not have access to Claude Fable. If your organization uses Workload Identity Federation instead of API keys, configure it for the same data retention workspace.