Skip to main content

Multi-Agent Studio

The Multi-Agent Studio (MAS) is the orchestration layer on top of the ITool registry. Where a single chat agent calls tools one at a time, MAS lets a director agent delegate to specialist agents that each own a slice of the toolset.

49+ specialists ship in 1.0, grouped under the CCGS roster (Component, Content, Gameplay, Systems).

Why specialists exist​

A single agent calling all 3,691 tools faces two problems:

  1. Planning saturation. Context windows fill up with tool descriptions. The LLM picks worse tools as the menu grows.
  2. Context contamination. A long debug session pollutes the agent's working memory with unrelated noise that hurts the next, unrelated task.

Specialists fix both. Each specialist has a filtered tool view — typically 80-200 tools relevant to its domain — and a clean context that starts fresh on each task.

The CCGS roster​

GroupSpecialists (selected)Tool families exposed
Componentgameobject-surgeon, prefab-doctor, transform-aligner, component-rewirergameobject_*, prefab_*, component_*
Contentmaterial-painter, mesh-optimizer, texture-importer, vroid-pipeline, clothing-rigger, pixel-cookmaterial_*, mesh_*, texture_*, vroid_*, clothing_*, pixel_*
Gameplayphysics-tuner, animation-author, audio-mixer, lighting-baker, nav-mesh-bakerphysics_*, animation_*, audio_*, lighting_*, nav_*
Systemsscript-author, build-engineer, test-runner, ci-doctor, headless-launcher, mcp-routerscript_*, build_*, tester_*, mcp_*

The director sees only the specialist roster, not the tools. When you ask "make my character wear this jacket", the director picks clothing-rigger. The director then hands off control with a focused brief — and clothing-rigger plans and executes inside its own context with its own ~150 tools.

Handoff workflow​

You ──▶ Director agent
              │
              │ (picks specialist + writes brief)
              ▼
        Specialist agent (e.g., clothing-rigger)
              │
              │ (calls 5-20 tools, returns artifact + report)
              ▼
        Director agent ──▶ You (synthesized result)

The transcript shows the handoff explicitly:

[director] Routing to clothing-rigger
  brief: "Rig 'Assets/Garments/Jacket.fbx' onto the avatar at 'Player' GameObject. ..."
[clothing-rigger] Loading 152 clothing_* and asset_* tools
[clothing-rigger] clothing_detect_garment_topology(...) → ok
[clothing-rigger] clothing_align_to_body(...) → ok
[clothing-rigger] clothing_bake_skin_weights(...) → ok
[clothing-rigger] Done. Returning artifact summary.
[director] Specialist returned. Synthesizing reply.

Authoring a new specialist​

A specialist is a Markdown file at Assets/Buril/Agents/<name>.md (or the user-scoped equivalent) with three sections:

# Specialist: <name>

## Tool view
asset_*
script_*
material_*

## System prompt
You are <name>. You specialize in <X>. Use tool families listed above only.
You return a one-paragraph report plus structured artifacts.

## Handoff schema
Inputs: { ... }
Outputs: { artifact, report }

On Editor reload, the registry scans for these files and adds them to the director's roster. There's no recompile required — drop the Markdown and the studio picks it up.

Customizing the director​

The default director is a general-purpose Claude-style agent. You can replace it by setting the BURIL_MAS_DIRECTOR model in Settings → Multi-Agent. Recommendations:

  • Anthropic Opus 4.7 — best planning, highest cost.
  • Anthropic Sonnet 4.7 — pragmatic default, balanced cost.
  • GPT-4o — competitive on planning, faster on simple routes.
  • Gemini 2.5 Pro — strong with long, document-heavy contexts.

Specialists can run on a smaller model than the director. The pattern that works in practice is Sonnet director / Haiku-class specialists — the director plans, the specialists execute.

Stopping a run​

The chat panel's Stop button cancels the current specialist call. If the specialist is mid-tool-call, the cancellation is cooperative — tools check a token at safe points and exit cleanly. Some long-running tools (e.g., build_player_batchmode) cannot be interrupted mid-build; they finish their current step, then abort.

When NOT to use MAS​

  • Single-tool calls (list scenes, play mode toggle) — the director overhead isn't worth it. Use the plain chat panel.
  • Heavily interactive work where you want to inspect every step — specialists batch 5-20 tools per run by design.
  • Tasks where you need precise control over which tool runs — call them directly.