AI - Building Agents

If you are looking to build an AI agent today, the landscape has split into two distinct paths: Custom Orchestration (building the engine yourself) and Managed Platforms (using pre-built environments like Google ADK or Claude’s emerging agentic suite).

The "Managed Platform" Path (High Speed, Enterprise Ready)

These options provide the "OS" for the agent. You don't build the loop; you just provide the instructions and the tools.

Google ADK (Vertex AI Agent Builder)

Google’s Agent Development Kit is designed for developers who want to build "Knowledge Agents" that are deeply integrated into the Google Cloud ecosystem.

The Building Block: "Playbooks." Instead of complex code, you define a Playbook—a natural language set of instructions that tells the agent how to handle specific scenarios.
Grounding: It is the best in class for "grounding" agents in reality. You can connect it to Google Search or your own BigQuery/Google Drive data with one click.
Why choose it: You need an agent that is safe, cited, and works with massive datasets (using Gemini’s 2-million-token window).

Claude Managed Agents (Anthropic)

Anthropic’s Managed Agents is their answer to the complexity of building agentic loops. It is designed to let you build powerful, multi-step agents without needing to manage the underlying infrastructure or state.

Managed Handoffs: It handles the transition between the model’s reasoning and the actual execution of tools. You no longer have to write the code that "waits" for a tool result; the platform manages the lifecycle of the task.
Native MCP Integration: Claude Managed Agents are designed to work seamlessly with the Model Context Protocol (MCP). This means you can "plug" a managed agent into a pre-built server (like your database or Slack) and it immediately knows how to use those tools.
Persistent Session State: The platform keeps track of the agent's progress, its "memory" of previous steps, and the status of its current task, allowing for long-running workflows that don't "forget" what they were doing.
Best For: High-reasoning tasks where the agent needs to "stop and think," use complex tools, or navigate a file system securely.

OpenAI Assistants API

The "original" managed agent service. It manages the "state" (memory) and the "threading" for you.

Why choose it: You want a simple, hosted solution where the AI remembers the conversation history and can search files without you managing a database.

The "Multi-Agent" Framework Path (High Control, Orchestration)

If you want to build a "team" of agents or have a very specific workflow that doesn't fit into a Google or Anthropic "Playbook," you use these frameworks.

LangGraph (by LangChain): This is for building "Custom Agents." It allows you to draw a flowchart of your agent's logic. If the agent fails at Step 3, LangGraph allows it to loop back to Step 1. It is the gold standard for high-precision, complex agents.
CrewAI: This is for "Team Orchestration." You define a "Manager Agent," a "Researcher Agent," and a "Writer Agent." You give them a task, and CrewAI manages the handoffs between them like a project manager.
AutoGen (Microsoft): Best for agents that need to write and execute their own code to solve math or engineering problems.

The "Developer-First" SDK Path

These are libraries for developers who want to write "Agentic Code" rather than "Agentic Prompts."

PydanticAI: A new, rising star. It brings strict "Type Safety" to agents. It ensures that when an agent uses a tool, the data it sends and receives is perfectly formatted, which is the #1 cause of agent failure in production.
Semantic Kernel: Microsoft’s engine for integrating AI into traditional apps (C#, Python, Java). It’s designed for industrial-grade apps where the agent is just one part of a larger software system.