Production-Ready Orchestration for AI Agents

AgentHelm is a lightweight Python framework for building AI agents with a focus on production-readiness. Built on DSPy, it provides plan-driven execution, full observability, and transactional safety guarantees.

import dspy
from agenthelm import ToolAgent, tool

@tool()
def search(query: str) -> str:
    """Search the web."""
    return f"Results for: {query}"


lm = dspy.LM("mistral/mistral-large-latest")
agent = ToolAgent(name="researcher", lm=lm, tools=[search])
result = agent.run("Find AI news")

# Generate a multi-step plan
agenthelm plan "Build a web scraper" -o plan.yaml

# Review and execute
agenthelm execute plan.yaml

# View execution traces
agenthelm traces list

# Filter and export
agenthelm traces filter --status failed
agenthelm traces export -o report.md -f md

# Connect to MCP servers
agenthelm mcp list-tools uvx mcp-server-time
agenthelm mcp run uvx mcp-server-time -t "What time is it?"

@tool(compensating_tool="reverse_payment")
def process_payment(amount: float) -> dict:
    """Process payment - rolls back on failure."""
    return payment_api.charge(amount)


@tool(requires_approval=True)
def delete_data(id: str) -> bool:
    """Dangerous operation - requires approval."""
    return db.delete(id)

pip install agenthelm

# Initialize configuration
agenthelm init

# Run a simple task
agenthelm run "What is 2+2?"

# Interactive chat
agenthelm chat

# Generate and execute plans
agenthelm plan "Build a calculator" -o plan.yaml
agenthelm execute plan.yaml

import dspy
from agenthelm import ToolAgent, tool, ExecutionTracer
from agenthelm.core.storage import SqliteStorage

@tool()
def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b


# Create agent with tracing
lm = dspy.LM("mistral/mistral-large-latest")
tracer = ExecutionTracer(storage=SqliteStorage("traces.db"))

agent = ToolAgent(
    name="calculator",
    lm=lm,
    tools=[add],
    tracer=tracer,
)

result = agent.run("What is 15 + 27?")
print(result.answer)  # "42"
print(f"Cost: ${result.total_cost_usd:.4f}")

Traces are automatically saved to ~/.agenthelm/traces.db:

# Custom storage path
agenthelm run "task" -s ./my_traces.db

# View traces
agenthelm traces list
agenthelm traces show 0

# Filter and export
agenthelm traces filter --tool search --status success
agenthelm traces export -o report.md -f md

# Start Jaeger
docker run -d -p 16686:16686 -p 4317:4317 jaegertracing/all-in-one

# Run with tracing
agenthelm run "task" --trace

# View in Jaeger UI: http://localhost:16686

graph TD
    A[Task] --> B[PlannerAgent]
    B --> C[Plan with Steps]
    C --> D[Orchestrator]
    D --> E[ToolAgent 1]
    D --> F[ToolAgent 2]
    E --> G[Tool Execution]
    F --> G
    G --> H[ExecutionTracer]
    H --> I[Storage Backend]

1. PlannerAgent - Generates multi-step plans from natural language
2. Orchestrator - Executes plans with parallel steps and Saga rollback
3. ToolAgent - ReAct reasoning loop with tool execution
4. ExecutionTracer - Captures every decision for observability
5. MCPClient - Connects to external MCP servers

📖 Full Documentation: https://hadywalied.github.io/agenthelm/

• Quick Start Guide
• CLI Reference
• API Reference
• Examples

• DSPy-native agents (ToolAgent, PlannerAgent)
• Plan-driven execution with YAML export
• Multi-agent orchestration with Saga pattern
• MCP protocol integration
• Full CLI with traces, chat, execute
• OpenTelemetry/Jaeger support
• Cost and token tracking

• Budget enforcement (tokens, time, I/O)
• Checkpointing for resumable workflows
• Policy engine for constraints
• Web dashboard for trace visualization

We welcome contributions! See CONTRIBUTING.md for guidelines.

• 🐛 Bug reports
• 💡 Feature requests
• 📖 Documentation improvements
• 🔧 Code contributions

MIT License - see LICENSE file for details.

• Documentation: hadywalied.github.io/agenthelm
• GitHub: github.com/hadywalied/agenthelm
• PyPI: pypi.org/project/agenthelm
• Author: Hady Walied | LinkedIn

⭐ Star us on GitHub if AgentHelm helps you build production-ready agents!