Beddel

Declarative YAML-based AI workflow engine for Python.

Define outcome-driven AI workflows in YAML — the engine handles adaptive execution with conditional branching, retry strategies, multi-provider LLM abstraction, and compositional primitives. YAML for the backbone, code escape hatches for complex logic.

steps:
  - id: greet
    primitive: llm
    config:
      model: gemini/gemini-2.0-flash
      prompt: "Say hello and share a fun fact about $input.topic"
      temperature: 0.7

Why Beddel

• Write workflows in YAML, not hundreds of lines of Python
• 7 built-in primitives cover most AI workflow patterns out of the box
• Multi-provider LLM support via LiteLLM (100+ providers)
• Adaptive execution: branching, retry with backoff, fallback, skip, delegate
• OpenTelemetry tracing with token usage tracking per step
• Lifecycle hooks for custom logging, metrics, and side effects
• Expose workflows as HTTP/SSE endpoints with one function call
• Hexagonal architecture — swap adapters without touching domain logic

Installation

# Core only (parser, resolver, executor — no external adapters)
pip install beddel

# With LLM adapters (LiteLLM, OpenTelemetry, httpx)
pip install beddel[adapters]

# With FastAPI integration (HTTP endpoints + SSE streaming)
pip install beddel[fastapi]

# With CLI (validate, run, serve workflows)
pip install beddel[cli]

# Everything
pip install beddel[all]

Requires Python 3.11+.

Quickstart

Get a workflow running in under 5 minutes.

1. Install

pip install beddel[adapters]

2. Set your API key

Get a free key from Google AI Studio:

export GEMINI_API_KEY="your-key-here"

3. Create a workflow

Save as workflow.yaml:

id: hello-world
name: Hello World
description: A minimal workflow that greets the user with a fun fact.

input_schema:
  type: object
  properties:
    topic:
      type: string
  required:
    - topic

steps:
  - id: greet
    primitive: llm
    config:
      model: gemini/gemini-2.0-flash
      prompt: "Say hello and share one fun fact about $input.topic"
      temperature: 0.7

4. Run it

import asyncio
from pathlib import Path

from beddel.adapters.litellm_adapter import LiteLLMAdapter
from beddel.domain.executor import WorkflowExecutor
from beddel.domain.parser import WorkflowParser
from beddel.domain.registry import PrimitiveRegistry
from beddel.primitives import register_builtins

async def main():
    workflow = WorkflowParser.parse(Path("workflow.yaml").read_text())

    registry = PrimitiveRegistry()
    register_builtins(registry)

    executor = WorkflowExecutor(registry, provider=LiteLLMAdapter())
    result = await executor.execute(workflow, inputs={"topic": "astronomy"})

    print(result["step_results"]["greet"]["content"])

asyncio.run(main())

python run_workflow.py

Model names use the LiteLLM format (provider/model). Avoid experimental (-exp) suffixes — they get retired without notice.

Features

Adaptive Core Engine (Epic 1)

The foundation. Parses YAML workflows, resolves variables, and executes steps with adaptive control flow.

YAML Parser — Secure loading via yaml.safe_load() with Pydantic 2.x validation. Supports workflow metadata, step definitions, variable references, conditional expressions, and execution strategy declarations.

Variable Resolver — Extensible namespace system with three built-in namespaces and a registration mechanism for custom ones:

prompt: "Tell me about $input.topic"           # Runtime inputs
prompt: "Expand on $stepResult.step1.content"   # Previous step outputs
prompt: "Using key $env.API_KEY"                # Environment variables

# Register custom namespaces
resolver.register_namespace("memory", my_memory_handler)

Adaptive Workflow Executor — Sequential execution with step-level conditional branching (if/then/else), configurable execution strategies per step, and step-level timeout support. The executor evaluates conditions and adapts flow — not a pure sequential dispatcher.

Execution Strategies — Five strategies per step, with exponential backoff and jitter for retries:

steps:
  - id: risky-call
    primitive: llm
    config:
      model: gemini/gemini-2.0-flash
      prompt: "Generate content about $input.topic"
    execution_strategy:
      type: retry
      retry:
        max_attempts: 3
        backoff_base: 2.0

Strategy	Behavior
fail	Stop workflow on error (default)
skip	Log error, continue to next step
retry	Retry with exponential backoff and jitter
fallback	Execute an alternative step on failure
delegate	Delegate error recovery to agent judgment

Primitive Registry — Instance-based registration with contract validation:

from beddel.domain.ports import IPrimitive
from beddel.domain.registry import PrimitiveRegistry

registry = PrimitiveRegistry()

class MyPrimitive(IPrimitive):
    async def execute(self, config, context):
        return {"result": "custom logic here"}

registry.register("my-custom-primitive", MyPrimitive())

Or use the @primitive decorator for module-level registration:

from beddel.domain.ports import IPrimitive
from beddel.domain.registry import primitive

@primitive("my-custom-primitive")
class MyPrimitive(IPrimitive):
    async def execute(self, config, context):
        return {"result": "custom logic here"}

LiteLLM Adapter — Multi-provider LLM abstraction supporting OpenRouter, Google Gemini, AWS Bedrock, Anthropic, and all LiteLLM-supported providers. Explicit API key resolution from environment variables for resilience against upstream library changes.

Compositional Primitives (Epic 2)

Seven built-in primitives that compose into complex agent behaviors.

Primitive	Description
llm	Single-turn LLM invocation with streaming support
chat	Multi-turn conversation with message history and context windowing
output-generator	Template-based output rendering (JSON, Markdown, text)
guardrail	Input/output validation with 4 failure strategies
call-agent	Nested workflow invocation with depth tracking
tool	External function invocation (sync and async)
agent-exec	Unified agent adapter for external agent delegation

chat — Multi-turn conversations with automatic context windowing:

steps:
  - id: conversation
    primitive: chat
    config:
      model: gemini/gemini-2.0-flash
      system: "You are a helpful coding assistant."
      messages:
        - role: user
          content: "What is Python?"
        - role: assistant
          content: "$stepResult.prev.content"
        - role: user
          content: "Tell me more about async/await"
      max_messages: 50
      max_context_tokens: 4000

guardrail — Validate LLM outputs with four failure strategies:

steps:
  - id: validate
    primitive: guardrail
    config:
      data: "$stepResult.generate.content"
      schema:
        fields:
          name: { type: str }
          age: { type: int }
      strategy: correct  # raise | return_errors | correct | delegate

Strategy	Behavior	LLM Required
raise	Hard fail with validation errors	No
return_errors	Soft fail — returns errors alongside data	No
correct	JSON repair (parse → strip markdown fences → retry)	No
delegate	Ask LLM to fix validation errors, retry up to N times	Yes

call-agent — Compose workflows by nesting them:

steps:
  - id: delegate
    primitive: call-agent
    config:
      workflow: summarizer-workflow
      inputs:
        text: "$stepResult.extract.content"
      max_depth: 5

tool — Invoke registered functions (sync or async):

steps:
  - id: search
    primitive: tool
    config:
      tool: web_search
      arguments:
        query: "$input.question"

# Register tools before execution
tool_registry = {
    "web_search": my_search_function,
    "calculate": my_calc_function,
}

Observability & Integration (Epic 3)

Production-grade observability and framework integration.

OpenTelemetry Tracing — Opt-in tracing with three nesting levels and token usage tracking:

from beddel.adapters.otel_adapter import OpenTelemetryAdapter

tracer = OpenTelemetryAdapter(service_name="my-app")
executor = WorkflowExecutor(registry, provider=adapter, tracer=tracer)

Spans generated:

• beddel.workflow — workflow-level span with beddel.workflow_id
• beddel.step.{step_id} — step-level span with token usage (gen_ai.usage.*)
• beddel.primitive.{name} — primitive-level span with model and provider attributes

Zero overhead when tracing is disabled (all calls gated behind if tracer is not None).

Lifecycle Hooks — Granular event system for custom logging, metrics, or side effects:

from beddel.domain.ports import ILifecycleHook

class MyHook(ILifecycleHook):
    async def on_workflow_start(self, workflow_id, inputs):
        print(f"Starting {workflow_id}")

    async def on_step_end(self, step_id, primitive, result):
        print(f"Step {step_id} completed")

    async def on_error(self, step_id, error):
        print(f"Error in {step_id}: {error}")

executor = WorkflowExecutor(registry, provider=adapter, hooks=[MyHook()])

Events: on_workflow_start, on_workflow_end, on_step_start, on_step_end, on_error, on_retry. Hook failures are silently caught — a misbehaving hook never breaks workflow execution.

FastAPI Integration — Expose workflows as HTTP/SSE endpoints with one function call:

from fastapi import FastAPI
from beddel.integrations.fastapi import create_beddel_handler

app = FastAPI()
router = create_beddel_handler(workflow)  # auto-creates provider + registry
app.include_router(router)

The handler streams workflow execution via Server-Sent Events (W3C-compliant). Clients receive real-time events: WORKFLOW_START, STEP_START, STEP_END, WORKFLOW_END.

pip install beddel[fastapi]
beddel serve -w workflow.yaml --port 8000

Endpoints:

• POST /workflows/{id} — Execute workflow (SSE response)
• GET /health — Health check

CLI

Beddel includes a command-line interface for validating, running, and serving workflows.

pip install beddel[cli]

Validate a workflow

beddel validate workflow.yaml

Output:

OK: hello-world
  name: Hello World
  steps: 1
  primitives: llm

Run a workflow

beddel run workflow.yaml --input topic=astronomy

Machine-readable output:

beddel run workflow.yaml --input topic=astronomy --json-output

List primitives

beddel list-primitives

Start the server

beddel serve -w workflow.yaml --port 8000
beddel serve -w flow1.yaml -w flow2.yaml --port 8000

Version

beddel version

OpenClaw Integration

Beddel works as an OpenClaw agent skill. After installing with pip install beddel[cli], the beddel command is available for any OpenClaw agent to create, validate, and execute AI workflows.

Practical examples

Agent that validates workflows before execution:

An OpenClaw agent can validate YAML files authored by users or other agents, catching schema errors before runtime:

openclaw agent --message "Validate my workflow at ./flows/pipeline.yaml" \
  --agent main

The agent calls beddel validate ./flows/pipeline.yaml and reports any issues.

Agent-driven workflow execution with dynamic inputs:

An OpenClaw agent can run Beddel workflows as part of a larger task, passing context-dependent inputs:

openclaw agent --message "Run the summarizer workflow for the topic 'quantum computing'" \
  --agent main

The agent calls beddel run summarizer.yaml --input topic="quantum computing" and processes the result.

Serving workflows for dashboard integration:

An OpenClaw agent can start the Beddel server to expose workflows as HTTP endpoints, enabling integration with dashboards or other services:

openclaw agent --message "Start the beddel server with all workflows in ./flows/" \
  --agent main

The agent discovers YAML files and runs beddel serve -w flow1.yaml -w flow2.yaml --port 8000.

Multi-agent pipeline with Beddel as the execution engine:

In a multi-agent setup, one agent (e.g., an architect) designs the workflow YAML, another (e.g., a QA agent) validates it, and a third executes it:

Architect agent → writes workflow.yaml
QA agent        → beddel validate workflow.yaml
Executor agent  → beddel run workflow.yaml --input topic=security --json-output

See SKILL.md for the full skill manifest and OpenClaw metadata.

Architecture

Beddel follows Hexagonal Architecture (Ports & Adapters). The domain core never imports from adapters or integrations — all external dependencies flow through port interfaces.

┌─────────────────────────────────────────────┐
│              Integrations                    │
│         FastAPI  ·  SSE Streaming            │
├─────────────────────────────────────────────┤
│               Adapters                       │
│    LiteLLM  ·  OpenTelemetry  ·  Hooks       │
├─────────────────────────────────────────────┤
│            Compositional Primitives          │
│  llm · chat · output · guardrail · tool · …  │
├─────────────────────────────────────────────┤
│              Domain Core                     │
│  Parser · Resolver · Executor · Registry     │
│  Models · Ports (interfaces)                 │
└─────────────────────────────────────────────┘

Development Setup

git clone https://github.com/botanarede/beddel-py.git
cd beddel/src/beddel-py
pip install -e ".[dev]"

Run all quality gates (recommended):

bash scripts/run-gates.sh

Or individually:

# Tests
python -Wd -m pytest

# Lint + format
ruff check .
ruff format .

# Type check
mypy src/

The -Wd flag turns DeprecationWarning into errors, catching deprecated API usage early.

Roadmap

Epics 1–3 (Adaptive Core, Compositional Primitives, Observability & Integration) are complete. Upcoming:

• Epic 4 — Adaptive Execution Patterns: reflection loops, parallel execution, circuit breaker, goal-oriented execution, MCP-native tool integration
• Epic 5 — Agent Autonomy & Safety: human-in-the-loop, model tier selection, PII tokenization, state persistence, cost controls

Contributing

Contributions are welcome. Open an issue to discuss before submitting a PR. Guidelines will be documented as the project matures.

Newsletter

License

MIT