Standard orchestration is weak. Static hierarchies are boring.
In ORC, leadership is earned in the Arena.
pip install orc-arenaimportasynciofromorcimportTheArena, Warrior, Elder# Create your warriorsgrog=Warrior(
name="Grog",
llm_client="gpt-4o", # Standard AI modelsystem_prompt="You are a senior backend dev", # Standard agent promptcapabilities=["code_review", "debugging"],
domains=["backend", "python"],
)
thrall=Warrior(
name="Thrall",
llm_client="claude-sonnet-4-20250514",
system_prompt="You are an infrastructure architect",
capabilities=["system_design", "scaling"],
domains=["backend", "infrastructure"], # Overlaps with Grog!
)
# The Elder judges all combatelder=Elder(judge=MetricsJudge())
# Enter the Arenaarena=TheArena(warriors=[grog, thrall], elder=elder)
result=awaitarena.battle("Optimize the database connection pooling")
print(f"Winner: {result.winner}")The wrapper is ORC-themed, but the arguments are standard AI concepts.
Warrior= Agent.Elder= Judge.TheArena= Orchestrator.
ORC (Orchestration by Ruthless Competition) is a multi-agent framework where AI agents compete for leadership through trials.
Unlike traditional orchestrators where a static "manager" routes tasks forever, ORC uses competitive dynamics:
- A task enters The Arena
- Warriors (agents) that claim the domain compete
- The Elder (judge) evaluates the combatants
- The winner becomes The Warchief — the leader for that domain
- The Warchief holds power until successfully challenged
It is an ever-changing, competition-based orchestration system.
┌─────────────────────────────────────────────────────────────────┐
│ THE ARENA │
│ │
│ ┌───────────┐ challenges ┌───────────┐ │
│ │ Warrior A │ ──────────────────> │ Warrior B │ │
│ │(Contender)│ │ (WARCHIEF)│ │
│ └───────────┘ └───────────┘ │
│ │ │ │
│ │ TRIAL BY TASK │ │
│ │ ┌───────────────────────┐ │ │
│ └───>│ Same task, both │<───┘ │
│ │ attempt solution │ │
│ └──────────┬────────────┘ │
│ │ │
│ v │
│ ┌───────────────────────┐ │
│ │ THE ELDER │ │
│ │ Evaluates quality │ │
│ └──────────┬────────────┘ │
│ │ │
│ Winner becomes / stays WARCHIEF │
└─────────────────────────────────────────────────────────────────┘
Once the Elder declares a victor, the winning Warrior is elevated to Warchief. They hold domain leadership until defeated.
- Dynamic Leadership — No hard-coded orchestrators. The best agent for the task takes command.
- Continuous Improvement — Agents must defend their position. Complacency means dethronement.
- Reputation System — Track agent performance across domains over time.
- Forced Rotation — Even dominant Warchiefs are rotated after too many consecutive defenses.
# Check who rules each domainwarchief=arena.get_warchief("backend")
print(f"Backend Warchief: {warchief}")
# Get the full leaderboardleaderboard=arena.get_leaderboard("backend", limit=5)
forentryinleaderboard:
crown="👑"ifentry["is_warlord"] else" "print(f" {crown}{entry['agent']}: rep={entry['reputation']:.2f}")Elders evaluate trial outcomes. Three built-in options:
fromorcimportLLMJudge, MetricsJudge, ConsensusJudge# LLM-based — an AI judges the AIelder=Elder(judge=LLMJudge(llm, criteria=["accuracy", "completeness", "efficiency"]))
# Metrics-based — cold, hard numberselder=Elder(judge=MetricsJudge(weights={"accuracy": 0.5, "latency": 0.3, "cost": 0.2}))
# Consensus — multiple judges voteelder=Elder(judge=ConsensusJudge([judge1, judge2, judge3]))Warriors can use different strategies for when to challenge the Warchief:
fromorcimportAlwaysChallenge, ReputationBased, CooldownStrategy, SpecialistStrategy# Berserker — always challengesgrog.challenge_strategy=AlwaysChallenge()
# Calculating — only challenges if reputation is higherthrall.challenge_strategy=ReputationBased(threshold=0.1)
# Patient — waits after losses, exponential backoffsylvanas.challenge_strategy=CooldownStrategy(base_cooldown=60)
# Specialist — only challenges in specific domainsgazlowe.challenge_strategy=SpecialistStrategy(specialties=["engineering"])| Traditional Orchestration | ORC |
|---|---|
| Central coordinator decides | Leadership emerges from competition |
| Static role assignment | Dynamic, earned leadership |
| Single point of failure | Any agent can lead |
| No quality pressure | Continuous improvement through trials |
| One agent does everything | Best agent for each domain rises |
- Agent A/B Testing — Compare agent implementations head-to-head on real tasks
- Model Evaluation — Pit GPT-4o vs Claude vs local models, get a leaderboard
- Self-Optimizing Systems — The best agent for each domain naturally rises to the top
- Research — Study emergent hierarchies in multi-agent systems
ORC provides two ways to use it:
fromorcimportTheArena, Warrior, Elder, Warchiefgrog=Warrior(name="Grog", llm_client="gpt-4o", system_prompt="...")
elder=Elder(judge=MetricsJudge())
arena=TheArena(warriors=[grog], elder=elder)
result=awaitarena.battle("task")fromorcimportArena, ArenaConfig, MetricsJudgearena=Arena(
agents=[my_agent_1, my_agent_2],
judge=MetricsJudge(),
config=ArenaConfig(challenge_probability=0.3),
)
result=awaitarena.process("task")Same engine. Same performance. Pick your style.
# Core (no LLM dependencies)
pip install orc-arena
# With LLM support
pip install orc-arena[openai] # OpenAI
pip install orc-arena[anthropic] # Anthropic
pip install orc-arena[ollama] # Ollama (local)
pip install orc-arena[all] # Everythingdocker build -t orc .
docker run orcgit clone https://github.com/Lumi-node/ORC.git
cd orc
pip install -e ".[dev]"
pytest tests/ -v
# Run examples
python examples/quick_battle.py
python examples/full_campaign.pyORC is powered by dynabots-core — a zero-dependency protocol foundation for multi-agent systems. ORC is the first in a family of orchestration frameworks, each exploring a different paradigm for coordinating AI agents.
Apache 2.0 - See LICENSE for details.