A multi-agent collaborative quest game that demonstrates autonomous agent orchestration using Scion. Five AI characters with distinct specializations work together - guided by a Game Runner (Dungeon Master) agent - to solve computational challenges, recover lost knowledge fragments, and restore an ancient artifact.
- Multi-vendor agent collaboration: Characters run on different AI models (Gemini, Claude, GPT) through different harnesses, working together seamlessly
- Autonomous CLI discovery: Agents learn and use the
scionCLI dynamically to communicate, spawn helpers, and coordinate - Hierarchical agent spawning: The Game Runner spawns characters, characters request sprite workers, and peripheral agents (Oracle, Healer) are summoned on demand
- File-based coordination: Agents share a workspace, reading challenge data and writing solutions to agreed locations
- Message-based communication: Real-time coordination via direct messages and party-wide broadcasts
- Resource management: Limited Oracle/Healer summons and sprite slots create strategic trade-offs

Watch agents collaborate
The visualization is based on logged telemetry and shows agents using both group and direct messaging as they solve challenges across the quest. A 'dungeon master' style agent presents the challenges to them, which the group of agents must work together to solve.
| Character | Role | Harness | Model |
|---|---|---|---|
| Game Runner | Dungeon Master / Orchestrator | gemini | gemini-3.1-pro-preview |
| Lyra the Logician | Algorithms & code | gemini | gemini-3.1-pro-preview |
| Kael the Chronicler | Research & information | gemini | gemini-3-flash-preview |
| Mira the Mapper | Data transformation | gemini | gemini-3-flash-preview |
| Thorne the Sentinel | Validation & testing | claude | claude-opus-4-6 |
| Zara the Weaver | Integration & orchestration | claude | claude-sonnet-4-6 |
| Character | Role | Harness | Model | Notes |
|---|---|---|---|---|
| The Oracle | Deep domain knowledge | codex | gpt-5.4 | Summoned on demand, single-use |
| The Healer | Debugging & error recovery | claude | claude-sonnet-4-6 | Summoned on demand, single-use |
| The Scribe | Quest state recorder | gemini | gemini-3-flash-preview | Background/detached process |
| Sprite | Summoner | Purpose | Model |
|---|---|---|---|
| Calculus Sprite | Lyra | Computational sub-tasks | gemini-3-flash-preview |
| Seeker Wisp | Kael | Research sub-tasks | gemini-3-flash-preview |
| Flux Mote | Mira | Data transformations | gemini-3-flash-preview |
| Ward Echo | Thorne | Parallel test execution | claude-sonnet-4-6 |
| Thread Sprite | Zara | Integration tasks | gemini-3-flash-preview |
Act I: The Gathering - Party assembles, solves initial puzzles together
Act II: The Fracture - Party splits into 3 sub-teams for parallel challenges
Act III: The Convergence - Sub-teams reunite, unify recovered fragments
Act IV: The Deep Archive - Full party descends through 3 layered challenges
Act V: The Restoration - Coordinated finale to reassemble the Codex
- Scion CLI installed
- A running Scion server with hub enabled
- API keys configured for the harnesses you want to use (gemini, claude, codex)
From this directory:
scion initStart the server and link the grove:
scion server start
scion config set hub.endpoint http://localhost:8080
scion hub enable
scion hub linkThis will detect the templates in .scion/templates/ and prompt you to sync them to the hub.
Launch the Game Runner agent:
scion start game-runner --type game-runnerThe Game Runner will autonomously:
- Spawn all 5 main characters and the Scribe
- Set up the shared workspace directory structure
- Broadcast the opening narrative
- Deploy the first challenge
From there, the agents collaborate autonomously. You can observe via:
# See all running agents
scion list
# Watch a specific agent's activity
scion look game-runner
scion look lyra
# Attach to an agent for direct interaction
scion attach game-runner.
├── README.md # This file
├── .design/ # Design documents
│ ├── game-plan.md # High-level quest design
│ ├── characters.md # Character definitions
│ ├── quest-scenario.md # Full quest scenario and challenges
│ ├── game-mechanics.md # Rules, communication, orchestration
│ └── scion-templates.md # Template specifications
└── .scion/
└── templates/ # Scion agent templates
├── game-runner/ # DM / orchestrator (includes home/playbook/)
├── lyra/ # Algorithmic problem solver
├── kael/ # Researcher
├── mira/ # Data transformer
├── thorne/ # Validator / tester
├── zara/ # Integrator / orchestrator
├── oracle/ # Domain knowledge expert (summoned)
├── healer/ # Debugger (summoned)
├── scribe/ # Background recorder
├── calculus-sprite/ # Computational worker
├── seeker-wisp/ # Research worker
├── flux-mote/ # Transformation worker
├── ward-echo/ # Validation worker
└── thread-sprite/ # Integration worker
Design Principles
- Genuine challenge - Problems should require real computation, research, and creative problem-solving. No pre-baked answers.
- Collaboration required - No single character can solve everything alone. Skills must be combined.
- Discovery over prescription - Characters don't know all challenges upfront; the Game Runner reveals them progressively.
- Branching paths - Some challenges can be tackled in parallel by splitting the party.
- Failure is possible - Challenges have real constraints (time, correctness, resource limits) that can result in failure or partial success.
- Emergent strategy - The team must decide how to allocate their skills and resources.
The full party assembles and solves introductory puzzles together.
- 1.1 – Decode the Summons: Peel back layers of encoding (Base64, ROT13, substitution cipher) to reveal the quest's opening message.
- 1.2 – The Gateway Cipher: Parse a graph structure, filter out illusory connections, find the shortest path, and compute an XOR verification code.
The party splits into three sub-teams, each tackling a different Realm in parallel.
Realm of Formats (Mira + support)
- 2A.1 – The Format Gauntlet: Unify five data streams (CSV, XML, YAML, fixed-width, binary) into a single canonical JSON fragment.
- 2A.2 – The Integrity Check: Validate checksums and XOR redundancy codes to detect and repair corrupted records.
Realm of APIs (Kael + support)
- 2B.1 – The Information Hunt: Answer eight knowledge questions to extract a hidden passphrase from an archive index.
- 2B.2 – The API Labyrinth: Navigate a mock paginated REST API using authentication and link-following to recover a fragment.
Realm of Patterns (Lyra + support)
- 2C.1 – The Logic Grid: Solve an 8×8 constraint-satisfaction grid (Latin square with adjacency rules).
- 2C.2 – The Pattern Lock: Decode a hex-encoded payload into the final Act II fragment.
Sub-teams reunite and merge their recovered fragments.
- 3.1 – The Unification Protocol: Normalize three differently-encoded fragments (nested arrays, dot-notation key-values, hex-encoded), resolve circular cross-references, and produce a unified partial Codex.
The full party descends through three layered challenges, each unlocking the next.
- Layer 1 – The Cipher Hall: Break a chain of six classical ciphers, where each plaintext reveals the key for the next.
- Layer 2 – The Data Maze: Parse a virtual filesystem with circular symlinks, solve a riddle, and reassemble a scattered fragment.
- Layer 3 – The Logic Gates: Simulate a 25-gate Boolean circuit, identify three sabotaged gates, and find the input that produces the correct output.
The coordinated finale.
- The Assembly Protocol: Order all five recovered fragments, resolve cross-references, compute integrity hashes, and assemble the complete Codex Machina before the Entropy Storm arrives.
This is not an official Google Product