choreoai 0.0.4

Build production-grade multi-agent AI systems — compose, orchestrate, and observe autonomous agents.

Build production-grade multi-agent AI systems.
Compose, orchestrate, and observe autonomous agents — reliably.

---

Choreo is an open-source framework for building multi-agent AI systems that survive production. Most agent frameworks make the happy path easy and leave you to discover the hard parts — hallucinations, runaway loops, lost state, unbounded cost — in incident review. Choreo makes those the default things you control: retries, guardrails, budgets, evals, and tracing are first-class, not add-ons.

Think of it as the layer between "a clever prompt" and "a system you can put on-call."

Status: pre-alpha, building in public. The APIs shown below are the target developer experience and are still landing — follow the roadmap. ⭐ Star the repo to track progress; issues and PRs are welcome.

Why Choreo

• Reliability is the product. Every run is bounded by step, token, time, and cost budgets. Nothing loops forever; nothing surprises your bill.
• Guardrails against bad states. Output/schema validation and grounding checks catch hallucinations before they propagate, with defined recovery paths (retry / replan / fall back / escalate).
• One base agent, thin subclasses. Cross-cutting concerns (LLM client, tool loop, retries, context assembly, tracing) live once in BaseAgent; your agents override only what's unique.
• Orchestration built in. Sequence, parallelize, or supervise workers — with shared state threaded cleanly through the run.
• Evals as a gate, not an afterthought. Per-agent, per-tool, and end-to-end checks so a prompt or model change can't silently regress.
• Observable by default. Full traces across agents and tool calls, plus metrics: latency, cost, retry rate, guardrail-trip rate, task success.

Install

# From PyPI (coming soon)
pip install choreoai

# From source (works today — installs the CLI and package skeleton)
pip install git+https://github.com/choreo-ai/choreoai

choreo            # confirm the install

Quickstart

Illustrative — this is the API Choreo is being designed around.

from choreo import BaseAgent, Orchestrator, budget

class Researcher(BaseAgent):
    system = "Find and summarize sources for the given topic."
    tools = [web_search]

fleet = Orchestrator(
    plan=Planner(),
    workers=[Researcher(), Researcher()],   # run in parallel
    reduce=Synthesizer(),
    review=Critic(),                         # gates the final answer
    budget=budget(steps=20, usd=0.50),       # hard cap on every run
)

result = fleet.run("Compare vector databases for RAG")

Roadmap

• Core — BaseAgent, tool loop, retries, tracing.
• Orchestration — planner + workers + synthesizer on a real research-assistant task.
• Reliability — budgets, guardrails, grounding checks, recovery paths.
• Evals & observability — eval harness, a critic agent, full-run traces.
• choreo CLI — init, run, trace.
• Docs site + first release to PyPI.

Contributing

Choreo is open source (MIT) and early — the best time to shape it. Bug reports, ideas, and PRs are all welcome. Open an issue to discuss a change before large PRs. Dev setup:

git clone https://github.com/choreo-ai/choreoai
cd choreo
pip install -e ".[dev]"
pytest

License

MIT © 2026 Karthik Reddy Basupally

_{Choreo · multi-agent systems, in production · github.com/choreo-ai/choreoai}