# Agent settings

> Settings › Agent in detail - sub-agent models, playtest verification, Blender control, thinking and transparency options, optimization mode, context compression, minified code and the maximum context window.

Open **Settings › Agent**. Changes apply to your account and are used by every new run.

## Agent & sub-agents
Big jobs are split across specialist [sub-agents](https://rogenie.ai/docs/sub-agents). Here you pick the model each one runs on.

- **Orchestrator model**: plans the work and assigns the sub-agents.
- **Specialist sub-agents**: Builder, Scripter, Debugger, UI Designer, VFX Artist, Asset & Audio Finder, Animator, 3D modelling (Blender), Verifier and Researcher.

**Same as chat** runs a role on the model you picked for the chat and sends no override. By default the Debugger, Asset & Audio Finder, Verifier and Researcher run on **Auto (intelligent & cheap)**; the other roles follow the chat, the Animator and 3D modelling (Blender) included. That sub-agent also sets up and checks its icon renders (Blender draws the image) on the model you pick for it, and that role works only while [Blender](https://rogenie.ai/docs/blender#blender-sub-agent) is connected. The status line says how many roles run on a model of their own.

### Playtest verification

RoGenie runs your game to verify that changes actually work. Turn **Playtest verification** off to skip runtime playtesting and save tokens. Static checks (compile, validation, scans) still run. See [Playtesting and verification](https://rogenie.ai/docs/playtesting).

### Blender control
**Let RoGenie control Blender** is on by default. With the [RoGenie Bridge](https://rogenie.ai/docs/bridge) running and Blender open on the same PC, RoGenie can model, rig and animate in your Blender; its Blender tools and the 3D modelling (Blender) sub-agent appear only while Blender is connected. The Animator uses Blender while it is connected and works in Studio otherwise. Turn it off to stop RoGenie using Blender even when it is open. See [Blender](https://rogenie.ai/docs/blender).

## Transparency

These controls show the thinking summaries a provider exposes (Claude thinking blocks, Gemini thought summaries) and rich tool-call logs. They never expose hidden raw chain-of-thought.

| Setting | What it does |
|---|---|
| **Enable thinking summaries** | Shows models' thought summaries live in the work log. |
| **Enable Claude extended thinking (when supported)** | Lets Claude models think before answering. |
| **Adaptive thinking (Opus 4.6 / Sonnet 4.6)** | Claude decides how deeply to think for each step. Recommended for those models; older models use a fixed budget. |
| **Claude thinking budget ratio (0.1–0.9)** | The fixed budget older Claude models get, as a share of the output. |
| **Enable Gemini thought summaries** | Shows Gemini's thought summaries. |
| **Enable working panel (tool calls + results)** | Shows the [work log](https://rogenie.ai/docs/watching-a-run#the-work-log) of tool calls and results under each reply. |

## Optimization mode

Every request re-sends the whole conversation, so how the agent paces its work changes the total cost.

| Mode | Pacing |
|---|---|
| **Efficient** (recommended) | Fewer, bigger replies. The lowest total token cost. |
| **Auto** | Quick, cheap steps first, then bigger plans once it has context. |
| **Granular** | More, smaller replies. Uses more total tokens; pick it only if you want the agent to check results after every small step. |
| **Model's choice** | No pacing guidance at all; the model decides how to split up the work. |

## Context and savings

RoGenie's "up to 11× more usage" comes from three levers. The first, discounted model rates, is automatic. The other two live here. See [How RoGenie saves tokens](https://rogenie.ai/docs/savings).

| Setting | What it does |
|---|---|
| **Smart context compression** | Structurally compresses large JSON tool outputs (scene reads, search results, property dumps) before they reach the model. Cuts token cost on long builds while preserving answers (about 97% fidelity, benchmark-backed). |
| **Minified code context** (experimental) | Shows the model a copy of your code with comments and extra whitespace removed but every line number kept, typically about a quarter fewer tokens, so edits land on the exact lines of your real scripts. Your scripts are never stored minified. |

### Maximum context window

The context window is a chat's memory (the [Chat memory](https://rogenie.ai/docs/chat-memory) ring by the message box): how much of the conversation RoGenie keeps in mind at once. It is **not** your usage limit.

RoGenie summarizes older turns once a chat reaches this many tokens, which keeps cost and latency down.

- **0** means auto: use the model's own window.
- Otherwise, enter at least **8,000** tokens.
- The value is always capped by the model. For example, 200K on a 1M-token model summarizes at 200K; 200K on a 100K-token model still caps at about 100K.

A lower value makes long chats cheaper per step, at the cost of summarizing more often.