​
Orkestra at a glance

  • SDK: Write small agents with typed I/O (Pydantic), retries, and tool access.
  • Orchestrator: Chain agents into playbooks/workflows (sequential/parallel/fallback) with shared state and policies (timeouts, retries, budgets).
  • Runtime: One command to expose agents/workflows as REST APIs for integration and testing.
  • Tools: Traces, logs, metrics, eval/replay, versioning, secrets, governance, and a no‑code builder.
You focus on agent logic. Orkestra handles deployment, wiring, reliability, and visibility.

​
Core concepts

  • Project: Namespace for agents, playbooks, policies, secrets, metrics.
  • Agent / AgentService: Stateless function handle(input) -> output packaged as a microservice.
  • Playbook (Workflow): Spec describing how steps execute (fan‑out/in, parallelism, timeouts, retries).
  • Invocation: One execution; produces a Trace (steps, inputs/outputs with redactions, latency, tokens, cost).
  • Policy: Versioned constraints (timeouts, retries, budgets, allowed tools/models).
  • Secret: Encrypted values injected at runtime.
  • Artifact: Files/blobs produced/consumed by steps, linkable from traces.

​
Get started

pip install orkestra-sdk

Quickstart

Install and run your first agent/workflow in minutes.

Overview

The Orkestra SDK, workflows, structured outputs, and deployment.

Agents

Define agents with typed inputs/outputs and generate responses.

Workflows

Chain agents with optional conditional logic (STOP/CONTINUE).

Structured outputs

Parse and validate model responses with Pydantic.

Server / APIs

Expose workflows as REST endpoints with Swagger and ReDoc.