Skip to content

Project overview

Agentic Runtime Platform is a monorepo for defining, running, observing, and evaluating multi-step AI workflows.

Last verified: 2026-10-18.

What the repository contains

Area Path Purpose
Shared tools repository root and tools/ Provider clients, benchmark helpers, cache, and shared errors
Runtime agentic-workflows-v2/ Workflow loader, execution engines, agents, tools, model routing, CLI, and server
Dashboard agentic-workflows-v2/ui/ Browser interface for workflows, models, runs, and evaluations
Evaluation agentic-v2-eval/ Offline scoring, metrics, evaluators, runners, and reports
Cross-package tests tests/e2e/ Verify public behavior across package boundaries

The repository root installs as the agentic-tools Python package. agentic-workflows-v2 and agentic-v2-eval are separate workspace members.

Main capabilities

Workflow runtime

  • Load typed workflow definitions from YAML.
  • Validate inputs, dependencies, conditions, and graph structure.
  • Run compatible workflows through either the native engine or LangGraph.
  • Route model-backed steps by tier, provider availability, and fallback policy.
  • Invoke approved tools and deterministic tier-0 steps.
  • Emit typed events and save typed run results.
  • Resume supported runs from checkpoints.

Server and dashboard

  • REST API for workflows, models, settings, runs, datasets, and evaluations.
  • WebSocket execution with replay support.
  • Server-sent events for chat and saved run streams.
  • Browser views for workflow editing, live execution, model configuration, run history, datasets, and evaluations.
  • API-key and optional OIDC authentication.
  • CORS, rate limiting, sanitization, audit logging, and tool boundaries.

These controls reduce risk but do not make every deployment secure by default. An exposed service still requires deliberate authentication, network, storage, secret, and tool-policy configuration.

Evaluation

There are two evaluation paths:

  • the runtime server scores saved runs for the API and dashboard;
  • agentic-v2-eval scores structured results offline with rubrics, metrics, batch runners, and JSON, Markdown, or HTML reports.

They exchange data rather than sharing an internal implementation.

Execution engines

Engine Role
native Dependency-light Kahn's-algorithm DAG executor; runs runtime-built DAG and Pipeline objects
langchain Compiles named YAML workflows to LangGraph; the default for agentic run and POST /api/run

The CLI defaults to langchain for named workflow runs. The native engine is not deprecated or replaced. Use agentic compare (and GET /api/adapters to see what is registered) when engine compatibility matters.

Technology

Area Main technology
Python packages Python 3.11+, hatchling, Pydantic v2
Runtime API FastAPI and Uvicorn
Workflow adapter LangGraph as an optional runtime extra
Dashboard React 19, TypeScript 7, Vite 8, TanStack Query, XYFlow
Tests pytest, Vitest, and Playwright
Documentation MkDocs Material
Telemetry Optional OpenTelemetry and OTLP
Containers Docker and Docker Compose
Automation GitHub Actions

The lockfile and package manifests are the source of truth for exact dependency versions.

Repository map

agentic-runtime-platform/
|-- pyproject.toml                 root agentic-tools package and workspace
|-- tools/                         shared Python source
|-- agentic-workflows-v2/
|   |-- agentic_v2/
|   |   |-- adapters/              engine registration
|   |   |-- agents/                agent implementations
|   |   |-- cli/                   Typer commands
|   |   |-- contracts/             public Pydantic wire models
|   |   |-- engine/                native execution
|   |   |-- langchain/             LangGraph adapter
|   |   |-- models/                routing and provider health
|   |   |-- server/                FastAPI, WebSocket, and SSE
|   |   |-- tools/                 runtime tools
|   |   `-- workflows/             loader and YAML definitions
|   |-- tests/                     runtime and server tests
|   `-- ui/                        React dashboard
|-- agentic-v2-eval/
|   |-- src/agentic_v2_eval/       offline evaluation package
|   `-- tests/
|-- tests/e2e/                     cross-package tests
|-- docs/                          maintained documentation and ADRs
|-- examples/                      small Python examples
|-- infra/                         deployment configuration
`-- otel/                          telemetry collector configuration

Current boundaries and limits

  • Provider-backed behavior requires current credentials and provider access.
  • AGENTIC_NO_LLM=1 is a deterministic development mode, not a provider integration test.
  • Native and LangGraph execution do not support every feature equally.
  • Some state is process-local even when Redis is configured.
  • Live provider evaluation is separate from the default unit-test path.

The maintained list is Known limitations.

Start here

Need Document
Install and run once Quick start
Learn the repository Contributor onboarding
Understand boundaries Architecture
Author workflow YAML Workflow authoring
Use the command line CLI reference
Call the server Runtime API
Configure providers and services Configuration
Develop and test Development guide
Deploy Deployment guide
Review decisions ADR index