isA Agent SDK

A complete AI Agent SDK for building intelligent agents with advanced features. Compatible with Claude Agent SDK patterns, with additional capabilities.

Features

• Claude Agent SDK Compatible - Familiar API patterns
• Streaming Messages - Real-time response streaming
• Built-in Tools - Read, Write, Edit, Bash, WebSearch, etc.
• MCP Integration - Model Context Protocol with offline fallback
• Human-in-the-Loop - Durable execution with checkpoints
• Skills System - Local-first skill loading with MCP fallback
• Project Config - .isa directory for project-specific settings
• Event Triggers - Proactive agent activation
• Hardware Service Clients - Device registry, telemetry and OTA integration through isA_user
• Multiple Execution Modes - Reactive, Collaborative, Proactive
• A2A Ready - Agent Card + JSON-RPC client/server adapters
• Multi-Agent Swarm - Dynamic agent handoffs with [HANDOFF:] directives
• DAG Task Execution - Dependency-ordered workflows with parallel wavefronts
• Production Resilience - Circuit breakers, bulkhead isolation, error classification, recovery playbooks
• Observability - OpenTelemetry tracing, structured error logging, Prometheus metrics
• Intent-Driven Routing - SenseNode classifies intent for automatic model tier selection
• Graceful Degradation - Nodes skip optional operations when services are degraded
• 3,500+ Tests - Comprehensive coverage including chaos tests

Multi-Agent Patterns

• Batch Swarm - Auto-decompose high-cardinality work, execute parallel DAG wavefronts, and synthesize the result.
• Multi-Agent Orchestration - Fixed teams, dynamic spawning, and tier selection.
• Swarm Orchestration - Handoff-driven agent collaboration.

Installation

pip install isa-agent-sdk

# With FastAPI server support
pip install isa-agent-sdk[server]

# With OpenTelemetry tracing
pip install isa-agent-sdk[observability]

Project Setup

Initialize a .isa directory for project-specific configuration:

mkdir -p .isa/skills

my-project/
├── .isa/
│   ├── config.json          # Project configuration
│   ├── settings.local.json  # MCP servers, permissions
│   └── skills/              # Local skills (loaded first)
│       └── my-skill/
│           └── SKILL.md
└── ...

Skills in .isa/skills/ are loaded before MCP, allowing project-specific overrides.

Quick Start

Basic Query

from isa_agent_sdk import query, ISAAgentOptions

# Simple usage
async for msg in query("Hello, world!"):
    print(msg.content)

# With options
async for msg in query(
    "Fix the bug in auth.py",
    options=ISAAgentOptions(
        allowed_tools=["Read", "Edit", "Bash"],
        execution_mode="collaborative"
    )
):
    if msg.is_text:
        print(msg.content, end="")
    elif msg.is_tool_use:
        print(f"[Tool: {msg.tool_name}]")

Human-in-the-Loop

from isa_agent_sdk import request_tool_permission, checkpoint

# Request permission before dangerous operations
authorized = await request_tool_permission(
    "delete_file",
    {"path": "important_data.txt"}
)
if authorized:
    # proceed with deletion
    pass

# Create checkpoints for durable execution
await checkpoint("about_to_deploy", {
    "version": "1.0.0",
    "environment": "production"
})

HTTP Client (for deployed apps)

from isa_agent_sdk import ISAAgent

client = ISAAgent(base_url="http://localhost:8000")

# Non-streaming
response = client.chat.create(
    message="Explain quantum computing",
    user_id="user123"
)
print(response.content)

# Streaming
for event in client.chat.stream(
    message="Write a Python function",
    user_id="user123"
):
    if event.is_content:
        print(event.content, end="")

Building a FastAPI Agent Service

from fastapi import FastAPI
from isa_agent_sdk import query, ISAAgentOptions

app = FastAPI()

@app.post("/query")
async def agent_query(prompt: str):
    responses = []
    async for msg in query(prompt):
        if msg.is_text:
            responses.append(msg.content)
    return {"response": "".join(responses)}

A2A (Agent-to-Agent) Integration

from fastapi import FastAPI
from isa_agent_sdk import (
    A2AAgentCard,
    A2AClient,
    A2AServerAdapter,
    register_a2a_fastapi_routes,
    build_auth_service_token_validator,
)

# Build agent card
card = A2AAgentCard(
    name="isA Agent",
    url="https://agent.example.com/a2a",
    token_url="https://auth.example.com/oauth/token",
).to_dict()

# Client call to remote A2A agent
client = A2AClient("https://remote-agent.example.com")
response = await client.send_message("https://remote-agent.example.com/a2a", "Hello from isA")

# Server adapter maps A2A JSON-RPC -> isa_agent_sdk ask/query
adapter = A2AServerAdapter()
rpc_result = await adapter.handle_rpc({
    "jsonrpc": "2.0",
    "id": "1",
    "method": "message/send",
    "params": {"message": {"parts": [{"text": "Summarize this repo"}]}}
})

# Mount A2A endpoints in FastAPI
app = FastAPI()
register_a2a_fastapi_routes(
    app,
    adapter=adapter,
    agent_card=card,
    rpc_path="/a2a",
    auth_validator=build_auth_service_token_validator(
        "http://localhost:8201",
        required_scopes=["a2a.invoke"],
    ),
)

Execution Modes

• Reactive - Responds to explicit requests
• Collaborative - Checkpoints for human approval
• Proactive - Anticipates needs, suggests actions

Documentation

• SDK documentation index
• Closed-loop runtime contracts

License

MIT