Scout

Scout is a lightweight model role for grunt work. When an Agent needs to do a lot of read-only exploration (searching a codebase, mapping out files, gathering web context), running every step on its expensive main model is slow and wasteful. Scout lets the Agent hand that exploration to a small, fast, cheap model and wait for a digest, the same way a developer using a coding assistant offloads heavy file-reading to a cheaper sub-agent instead of burning premium steps on it.

What scout actually is

There are two related things called “scout”: a tool an Agent calls, and a model role that tool runs on.

The scout tool

scout is a native tool available to both main Agents and sub-Agents. When an Agent calls it with a self-contained brief, the tool:

1. Resolves a cheap “scout” model (see the resolution chain below).
2. Spawns an await sub-task on that model, using the read-only scout toolbox (grep, read_file, list_directory, web_search, browse_url, extract_links). The scout toolbox has no write tools and no spawn or scout tools, so a scout is always a leaf: it explores, it cannot mutate anything, and it cannot delegate further.
3. Blocks the calling Agent until the scout returns its digest, which becomes the tool’s result and arrives as the Agent’s next message.

The scout sub-task has no view of the calling conversation, so the brief must be self-contained: what to find or read, which paths or queries to start from, and the shape of digest to report back. The tool accepts optional hints (focus paths and suggested searches) that are folded into the brief.

This is the only confirmed use of scout in the codebase: delegated, read-only exploration via the scout tool. It is not used for compacting, memory extraction, or message titling; those are separate model roles (see below).

Where the scout tool is available

The scout tool is included in the built-in code, research, and ops toolboxes, so any Agent with one of those (or the all toolbox, or a custom toolbox that lists scout) can delegate exploration. See Toolboxes for how grants work.

How the scout model is resolved

The scout model is resolved through a fallback chain, most specific first. The first tier that has a non-empty model wins:

1. Per-call override: an explicit model (plus its provider_id) passed to the scout tool for that call only.
2. Agent scout: the Agent’s own scoutModel / scoutProviderId.
3. Project scout: when the scout runs in a project context, the project’s scoutModel / scoutProviderId.
4. Global scout default: the platform-wide default (default_scout_model / default_scout_provider_id).
5. The Agent’s own main model: the safety net.

Because the chain ends at the Agent’s own model, scout is purely additive: on an install with no scout configuration at all, every scout simply runs on the calling Agent’s main model. Nothing breaks; you just do not get the cost savings until you point scout at a cheaper model.

Note

When you override the scout model on a single scout call, you must pass provider_id too. The same model name can be served by more than one provider, and Hivekeep will not guess which one you mean.

Configuring scout

Scout is configured through the UI, at three levels matching the resolution chain. (There is no environment variable for the scout default; it is stored as a platform setting.)

• Global default: in Settings → Models & services, set the Default Scout Model. A small, fast model is ideal here. When unset, scouts fall back to the calling Agent’s own model.
• Per Agent: in an Agent’s settings, set its Scout model. Leave it on inherit to fall back to the project scout default, then the global default, then the Agent’s own model.
• Per project: in a project’s settings, set its Default scout model to override the global default for tasks on that project’s tickets.

These can also be set programmatically. The global default is one of the model-bearing services handled by the set_default_model tool, alongside llm, embedding, image, compacting, and extraction:

set_default_model(service: "scout", model: "claude-haiku-4-6", provider_id: "anthropic")

Queenie can do this for you in conversation, and get_default_models reports the current scout default along with the rest.

When to use scout

Reach for scout when an Agent would otherwise spend many of its own steps reading and grepping before it can act. A good brief asks the scout to gather and summarize, then hands the digest back so the main Agent can reason on a compact result rather than dozens of raw file reads. Because the scout runs read-only and cannot write or spawn, it is safe to delegate freely.

Related pages

• Toolboxes: the scout toolbox the scout runs with, and the toolboxes that grant the scout tool.
• Choosing a Model: picking models for each role.
• Queenie, guided setup: can set your default scout model in conversation.