Agentic Runtime Platform¶

Agentic Runtime Platform runs YAML-defined AI workflows. Each workflow is a directed acyclic graph: steps declare their inputs, outputs, and dependencies; the runtime schedules ready steps, runs independent work concurrently, and prevents downstream steps from using failed or missing outputs.

The repository includes a Python runtime and CLI, a FastAPI server, a React dashboard, an evaluation package, and shared model utilities. It supports local development without provider credentials and real provider calls when credentials are configured.

Run the quick start Read the architecture Check known limitations

Quick start¶

Install the workspace:

git clone https://github.com/tafreeman/agentic-runtime-platform.git
cd agentic-runtime-platform
just setup

Create a JSON input file and run the deterministic workflow:

printf '{"input_text":"Hello World"}\n' > /tmp/agentic-input.json
AGENTIC_NO_LLM=1 agentic run test_deterministic \
  --input /tmp/agentic-input.json

PowerShell:

'{"input_text":"Hello World"}' |
  Set-Content -Encoding ascii .\agentic-input.json
$env:AGENTIC_NO_LLM = "1"
agentic run test_deterministic --input .\agentic-input.json

Expected result: Status: SUCCESS. The input option accepts a file path, not inline JSON.

AGENTIC_NO_LLM=1 replaces model calls with a fixed response. It verifies workflow structure, server integration, and event streaming. It does not verify response quality or structured model output.

Detailed quick start

Choose a starting point¶

How a run works¶

flowchart LR
    D["YAML definition"] --> L["Loader and validation"]
    L --> G["Step graph"]
    G --> E{"Execution adapter"}
    E -->|native| N["Native DAG executor"]
    E -->|langchain| LC["LangGraph adapter"]
    N --> R["Model router and tools"]
    LC --> R
    R --> O["Run result and events"]
    O --> U["CLI, API, and dashboard"]
    O --> V["Evaluation"]

Named YAML workflows default to the LangGraph adapter. Pass --adapter native for the built-in DAG engine. Runtime-generated DAG and Pipeline objects use the native engine. Both engines remain supported.

Main capabilities¶

Workflow execution¶

• YAML input and output contracts
• Dependency-aware scheduling
• Concurrent branches and fan-in
• Conditional steps and bounded loops
• Retries, timeouts, and failure propagation
• Run records plus SSE and WebSocket events

Model and tool control¶

• Capability-tier routing across configured providers
• Health tracking, fallback chains, and circuit breakers
• Explicit tool allowlists
• Approval checks for high-impact tools
• File path containment and outbound URL checks

Approval-required calls are denied when no approval provider is registered. This is safer than silently running the tool, but it also means a deployment must connect an approval provider before those calls can succeed.

Evaluation¶

• YAML-defined rubrics
• Objective metrics that run without a model
• Batch and streaming runners
• Optional LLM-as-judge scoring
• JSON, Markdown, and HTML reports

Repository numbers¶

These values are generated from the current source tree:

56 architecture decision records capture accepted, rejected, superseded, and proposed choices. The ADR index is the source of truth for their status.

The test count is a static count of Python test functions. It is not a claim that every test passed in the latest local run. Check the repository's CI workflow for current hosted results.

Development commands¶

Run these from the repository root:

just test
just docs
pre-commit run --all-files
npm --prefix agentic-workflows-v2/ui run build

The docs check validates selected file references and verifies that the numbers above match the source. The published site is also built with MkDocs strict mode in the documentation workflow.

Scope and limitations¶

The repository includes controls needed for serious development and testing, but deployment still requires operator decisions:

• configure API-key or OIDC authentication;
• place shared rate limiting at the edge for multiple replicas;
• choose durable stores for run records and replay data;
• register an approval provider for approval-required tools;
• apply network-level egress controls for high-assurance environments;
• run provider-specific integration tests for every configured provider.

Review Known limitations, Security hardening, and Deployment before exposing the service.