Coming Soon

Workflows that
think for themselves.

Orchestras are multi-step workflows where an AI Director coordinates tasks, humans make decisions in real-time, and everything runs as ephemeral containers on auto-provisioned infrastructure.

How it works

An orchestra, not a pipeline

orchestra = cc.create_orchestra("product-cleanup",
director="ai",
ai_config={"provider": "anthropic", "model": "claude-sonnet-4-20250514"},
first_agent_node="prc_scrape",
)
# The AI Director takes over:
# 1. Scraper runs → finds 150 products
# 2. AI decides: "too many, need validation" → triggers validator
# 3. Validator finds 30 invalid → AI asks human: "Delete 30 products?"
# 4. Human clicks [Delete expired] [Export CSV] [Ignore]
# 5. Human chose "Delete expired" → AI triggers delete process
# 6. AI posts to chat: "Cleanup complete. 25 deleted, 5 exported."
Orchestra

A group of related executions working toward a goal

Director

The decision-maker — code, AI, or human

"Musician" (AgentNode)

An autonomous node that does actual work and returns a result

Score

The full execution history with results and decisions

Capabilities

Everything an orchestra
needs to perform

🎼

Director Pattern

A Director (code or AI) coordinates AgentNodes (tasks). Each AgentNode does its work and returns a result. The Director decides what happens next.

🤖

AI Director

Plug in Claude, GPT, or Gemini as your director. The AI reads the full execution history and autonomously decides the next step — or asks a human when unsure.

💬

Real-Time Chat

Every orchestra has a live chat channel. Director, AgentNodes, and humans can post messages, share files, request help, and trigger actions.

🙋

Human-in-the-Loop

AgentNodes can present N choices to a human. Each choice links to a different next step. Not just approve/reject — real multi-path decisions.

📦

Container Execution

Run AgentNodes as ephemeral Docker containers on auto-provisioned Hetzner servers. 0.5 CPU per AgentNode, Swarm handles scheduling.

🔐

Workspace Secrets

Encrypted vault for API keys, credentials, and tokens. Injected as environment variables into AgentNodes. AES-256-GCM at rest.

📎

Artifacts & Results

AgentNodes store structured JSON results and upload files (S3/local). The next AgentNode reads the result — no shared database needed.

Timeout & Budget

Set a max duration, max AI calls, or max movements. Warnings at 80%, automatic failure at 100%. Pause and resume anytime.

🌐

5 Execution Methods

AgentNodes can run via HTTP, SSH, AWS SSM, Kubernetes Jobs, or Docker containers. Mix and match within the same orchestra.

Comparison

What no one else
combines

Feature CronControl Temporal Trigger.dev Airflow
Director decides next step at runtime
AI Director (Claude/GPT/Gemini)
Human-in-the-loop with N choices
Real-time chat between tasks
Language agnostic (any language)
Container execution (Docker)
Auto-provisioned infrastructure
HTTP/SSH/SSM/K8s execution
Open source (MIT)
No process stays alive waiting

See full comparison →

Want early access?

Orchestras are in active development. Get notified when they ship.