diff --git a/crates/settings_content/src/agent.rs b/crates/settings_content/src/agent.rs index b066bdaf298..4078f7aebb4 100644 --- a/crates/settings_content/src/agent.rs +++ b/crates/settings_content/src/agent.rs @@ -282,7 +282,9 @@ pub struct AgentSettingsContent { pub max_content_width: Option, /// The default model to use when creating new chats and for other features when a specific model is not specified. pub default_model: Option, - /// The model to use for subagents spawned via the `spawn_agent` tool. Defaults to the parent agent's model when not specified. + /// Model to use for subagents. Defaults to the parent thread's model and its + /// thinking, effort, and speed settings when not specified. If the configured + /// model is unavailable, the parent thread's model is used instead. pub subagent_model: Option, /// Favorite models to show at the top of the model selector. #[serde(default)] diff --git a/docs/src/ai/agent-settings.md b/docs/src/ai/agent-settings.md index 406695f0d8a..e92d93d6d8f 100644 --- a/docs/src/ai/agent-settings.md +++ b/docs/src/ai/agent-settings.md @@ -156,6 +156,50 @@ Zed supports feature-specific model settings for Inline Assistant, Git commit ge See [LLM Providers](./llm-providers.md) for model access, and [All Settings](../reference/all-settings.md) for the complete settings reference. +### Subagent Models {#subagent-models} + +By default, Zed Agent's subagents inherit the parent thread's model, thinking setting, reasoning effort, and speed. Set `agent.subagent_model` in `settings.json` to choose separate model, reasoning, and speed settings. You can override the model by asking the agent to use a specific model for a subagent. When that model matches the provider and model in `agent.subagent_model`, its `enable_thinking`, `effort`, and `speed` settings are preserved. This setting does not configure [External Agents](./external-agents.md). + +Configure this setting with a `provider` and `model`. To choose a reasoning effort for a model that supports it, also set `enable_thinking` and `effort`: + +```json [settings] +{ + "agent": { + "subagent_model": { + "effort": "low", + "enable_thinking": true, + "model": "claude-opus-4-6", + "provider": "anthropic" + } + } +} +``` + +Choose an `effort` value supported by the selected model. Omitting `effort` from this object does not inherit the parent's effort. It leaves effort to the selected provider and model's default behavior. Omitting `enable_thinking` defaults to `false`, although models that cannot disable thinking still use it. + +To find supported effort values: + +1. Choose the target provider and model from the model selector in a native Zed Agent thread, even if it is already selected, to save it as `agent.default_model`. +2. Enable thinking if it is off, then open the thinking effort menu. The "Change Thinking Effort" menu lists the model's supported levels. If the model offers no effort choices, omit `effort`. +3. Choose a level and read `agent.default_model.effort` in `settings.json` for the exact value to use in `agent.subagent_model.effort`. + +These controls change the current thread and saved default model settings, and can update a matching favorite model. Restore your previous selections afterward if needed. + +The optional `speed` field accepts `"fast"` or `"standard"`. It applies only to models that support fast mode. Omitting it from `agent.subagent_model` leaves speed unspecified rather than inheriting the parent's speed. + +| Model Selection | Settings Applied | +| --------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | +| No specific model requested, and `agent.subagent_model` is unset | Inherits the parent thread's current thinking, effort, and speed settings. | +| No specific model requested, and `agent.subagent_model` is configured | Uses `enable_thinking`, `effort`, and `speed` from `agent.subagent_model`. | +| Requested model matches the provider and model in `agent.subagent_model` | Uses `enable_thinking`, `effort`, and `speed` from `agent.subagent_model`. | +| Requested model differs from `agent.subagent_model`, or `agent.subagent_model` is unset | Uses matching `agent.default_model` settings. Without a match, enables thinking if supported, uses the model's default effort when available, and leaves speed unspecified. | + +When using matching `agent.default_model` settings, unsupported or omitted `effort` falls back to the selected model's default effort, when available. + +If the configured model is unavailable, subagents use the parent thread's model and its thinking, effort, and speed settings. + +The agent can only request a model for a subagent, not a separate effort or speed. Requesting the parent's own model is not the same as requesting no specific model: the table's rules for requested models apply instead of inheriting the parent's current settings. Subagents with a requested or configured model do not follow later changes to the parent's model, reasoning, or speed settings. Resuming a subagent keeps its existing model, reasoning, and speed settings. + ## Model Temperature {#model-temperature} Most Zed AI features use the selected model's default generation behavior. diff --git a/docs/src/ai/tools.md b/docs/src/ai/tools.md index 3b47da18798..ec43506c955 100644 --- a/docs/src/ai/tools.md +++ b/docs/src/ai/tools.md @@ -109,4 +109,6 @@ Loads instructions from an available [Skill](./skills.md) so the agent can follo Spawns a subagent with its own context window to perform a delegated task. Useful for running parallel investigations, completing self-contained tasks, or performing research where only the outcome matters. Each subagent has access to the same tools as the parent agent. +To configure a subagent's model, reasoning effort, and speed, see [Subagent Models](./agent-settings.md#subagent-models). + **Example:** While refactoring the authentication module, spawn a subagent to investigate how session tokens are validated elsewhere in the codebase. The parent agent continues its work and reviews the subagent's findings when it completes — keeping both context windows focused on a single task. diff --git a/docs/src/reference/all-settings.md b/docs/src/reference/all-settings.md index e15231f5b46..fc14b42f7e8 100644 --- a/docs/src/reference/all-settings.md +++ b/docs/src/reference/all-settings.md @@ -6062,6 +6062,11 @@ Visit [AI Quick Start](../ai/quick-start.md) under the AI section to learn more `boolean` values +### Subagent Model {#agent-subagent-model} + +- Description: The model, reasoning, and speed settings for new Zed Agent subagents. This setting does not configure external agents. See [Subagent Models](../ai/agent-settings.md#subagent-models). +- Setting: `agent.subagent_model` + ## Collaboration Panel - Description: Customizations for the collaboration panel.