Skip to content

Getting started

Three entry points, by audience:

Pick your path

1. The 60-second evaluator

You want to confirm the runtime works on your machine and produces sensible artifacts before committing real reading time.

  • Time: about 60 seconds after install completes
  • Credentials: none — AGENTIC_NO_LLM=1 runs the full DAG against deterministic placeholders
  • Output: a structured run record with step timings, tool calls, and a final artifact
  • Continue with: Quick start

2. The workflow author

You want to write your own multi-agent workflows. You have likely already seen the demo and want to understand the YAML grammar and step contracts.

  • Time: 15–30 minutes for the first hand-authored workflow
  • Credentials: still optional — AGENTIC_NO_LLM=1 lets you iterate on graph shape and contracts before any provider is wired
  • Output: a custom workflow definition that runs end-to-end
  • Continue with: First workflow, then the Workflow authoring guide

3. The platform integrator

You are wiring real providers, the FastAPI server, and the React dashboard into a longer-lived environment — staging, demo, or a customer trial.

  • Time: 1–2 hours including provider keys, OpenTelemetry collector, and the UI build
  • Credentials: at least one real LLM provider key (OpenAI, Anthropic, Gemini, Azure OpenAI, Azure Foundry, GitHub Models, Ollama, or local ONNX)
  • Output: a running server on port 8010 with the live UI on 5173
  • Continue with: Installation, then the Architecture overview

What you get

Same surface area regardless of path:

Artifact Where it lives Why it matters
agentic CLI Installed on $PATH after pip install -e Lists workflows, agents, tools; runs and validates workflows; ingests RAG documents
FastAPI server agentic-workflows-v2/agentic_v2/server/ REST + WebSocket + SSE surface for browser clients and long-lived integrations
React 19 dashboard agentic-workflows-v2/ui/ Live DAG visualizer with step-level streaming, run history, and rubric scores
Evaluation harness agentic-v2-eval/ Rubric runner with batch and async streaming modes, plus the LLM-as-judge protocol
Shared LLM client tools/ (agentic-tools package) Multi-provider client with caching, retries, and the canonical error taxonomy

Prerequisites

  • Python 3.11 or newer. The runtime exercises asyncio.TaskGroup and Self from typing — there is no fallback path for 3.10 or older.
  • Node 20+ and npm 10+ for the dashboard. The frontend uses Vite 8 and React 19, both of which require modern toolchains.
  • Git with the ability to fetch full history. The git-revision-date-localized plugin embeds last-modified dates into the documentation site, and CI runs with fetch-depth: 0 for that reason.

No Docker, Postgres, or cloud account needed to run the deterministic test workflow. Provider credentials matter only once you leave No-LLM mode.

Common starting commands

# Show every workflow and agent the runtime knows about
agentic list workflows
agentic list agents

# Validate a workflow definition without executing it
agentic validate code_review

# Compare native and LangGraph adapters on the same input
# (run from agentic-workflows-v2/; the fixture ships with the test suite)
agentic compare test_deterministic --input tests/fixtures/deterministic_input.json

If you get stuck, the Known Limitations page is an honest accounting of items that are shipped with caveats — start there before opening an issue.