The Ohtani vs. Freeman vs Mookie Trade-Off: Picking an AI Agent Framework
In my last blog post, I discussed the emergence of agentic AI in the federal government, including its use cases, as well as the security and infrastructure considerations that come with implementation.
To recap, agentic AI’s promise lies in its ability to enable agents that can plan, coordinate, and complete multi-step processes independently often in collaboration with other agents.
In this blog post, I will go a step further into technical implementation, including a review of several AI agent frameworks. I’ll explore the flexibility vs convenience trade-off and map which frameworks best suit certain government missions. And yes, I’ll be throwing in a few baseball analogies and jokes along the way, because why not? (1).
If we're going to talk about AI agents, we might as well make it fun.
Footnotes are at the bottom. Yes, I know, I put the jokes down there. Blame Bill Simmons.
What is an Agent Framework?
I know, you see the term “framework” in consultant-speak and roll your eyes, so I’ll compare it to baseball. Think of it this way: the LLM (e.g. GPT, Claude, Gemini, etc.) is the team’s star, the lifeblood of the roster. The framework is the coaching staff, setting the lineup, moving players around, and making sure the star is in a position to succeed.
In technical terms, the framework takes the LLM’s raw talent and channels it through a structure that lets agents coordinate, reason, and act within clear boundaries (2).
Every agent framework provides structure that allows agents to:
- Interpret the goals or instructions
- Plan and sequence actions
- Execute actions via APIs, databases, or other systems
- Reflect and adjust based on results
In an IT environment, an agent framework plays a role similar to Jenkins or Harness, which manage workflows. The difference is that while DevOps tools require human triggers to start processes, agent frameworks can drive workflows autonomously, enabling agents to make multi-step decisions without constant oversight.
The AI Agentic Framework Landscape
There is no single “right” framework for every situation. Choosing the best one is like building a baseball team roster.
You can sign a superstar like Shohei Ohtani, a unicorn that can do anything (3) including throwing 100+ mph and hitting baseballs into orbit, but that versatility requires careful management and scheduling to get the best performance (Open Source - High Flexibility, Low Convenience).
You can bring in a system player like Mookie Betts, an elite talent that can play for anyone and integrates into whatever role the system needs (4) (Enterprise - Moderate Flexibility, Moderate Convenience).
Finally, you can rely on a veteran like Freddie Freeman, who consistently performs well with minimal oversight, though with less room for customization or creative use (5) (Low Flexibility, High Convenience).
Open-Source Frameworks: Maximum Flexibility (Ohtani)
On the tech side, open-source frameworks (High Flexibility, Low Convenience) like LangChain, LangGraph, and CrewAI are your Shohei Ohtanis: they are flexible and powerful and give you full control over reasoning logic, agent memories (yes, agents have memories!), and guardrails, but you need to manage all the details yourself.
CrewAI shines when multiple agents need to coordinate or challenge each other. It’s perfect for research tasks, debates, or synthesis work, and can support “debate” or “adversarial synthesis” frameworks, where agents propose ideas and other agents take opposing positions.
Think of it like a well-functioning baseball front office at the trade deadline: one scout argues for trading for a starting pitcher, another pushes for a utility infielder, analytics weighs in with stats, and the General Manager steps in to make the final call (6).
LangChain/LangGraph is particularly good at connecting LLMs to structured data and managing complex reasoning chains. LangGraph’s graph-based design is especially powerful for relationship mapping and managing data dependencies.
Newer open-source frameworks like Pydantic Agents and Marvin prioritize reliability. These models use strict data templates or commands to ensure outputs are perfectly structured and consistent. This is an important feature for federal use cases like audit and compliance, where eliminating formatting errors and ensuring traceable data is important.
Enterprise Frameworks: The Middle Ground (Betts)
In the middle are enterprise frameworks, your Mookie Betts types, such as Microsoft Agent Framework, Google ADK, and AWS Bedrock AgentCore and Strands Agent SDK.
These are the Moderate Flexibility / Moderate Convenience tools that integrate directly into cloud environments with compliance and security built in. Agencies already working with these cloud providers may choose them because of ease of implementation and integration with existing enterprise workflows.
These frameworks are also helpful for classified environments and where FedRAMP High certification is paramount. Agencies investing in Zero Trust can also use these frameworks to strengthen data sovereignty.
Hosted Frameworks: Maximum Convenience (Freeman)
On the far end of flexibility-convenience are hosted frameworks, your Freddie Freemans, such as the Open AI Agent Framework/Agent Kit.
These frameworks are Low Flexibility / High Convenience because they are fully managed by the provider. The trade-off is that they are easy to implement, making them great for experimentation or low-risk use cases, but they offer little flexibility.
They are best used for public-facing or non-sensitive data.
The Key AI Agent Standard: Model Context Protocol (MCP)
Model Context Protocol is not a framework in and of itself, but it is an important standard for securely connecting AI agents to external tools. Examples of these tools include data scrapers, database tools, code execution tools, DevOps tools, search tools, messaging services, and API connectors. MCP is the connective piece that keeps agents aware of context and what they’re allowed to access. It ensures that when agents use APIs and pull data, they can do so consistently.
MCP also enforces security and guardrails through Role-Based Access Control (RBAC) for tools, ensuring an agent can only access the data/tools its mission and user privileges allow.
An MCP server is simply a packaged tool that channels its services through the Model Context Protocol standard. Instead of building every agent-tool integration from scratch, you can “plug in” an MCP server to instantly give an agent access to any of the tool skills.
There are several platforms that provide MCP servers for use, including mcp.so, smithery.ai, and glama.
Think of MCP as the dugout’s communication system: it keeps every player (agent) aware of the game situation (context) and ensures they only do what they’re authorized to do (permissions). Each MCP server is like a specialized player on your roster who is called in when you need that skill. You don’t need to train every agent from scratch, you just put the right player in the game in the right situation.
You just make sure they stay healthy because depth matters. (7).
Matching Agentic AI Frameworks to Real Life Use Cases in Government
In my previous post, I listed different use cases where agentic AI can help streamline operations and improve cost-effectiveness. Here's how agentic AI frameworks can be deployed for these missions:
Document Intelligence
Use Case: Researching contracts, policies, and recommendations
This use case is a great fit for LangChain/LangGraph and CrewAI. LangChain/LangGraph excels at connecting large language models (LLMs) to structured data sources like contracts and policies. This enables textual data extraction from multiple sources, relationship mapping, and semantic search.
Similarly, CrewAI can support multi-agent collaboration for complex research tasks, including creating legal research and debate agents that challenge each other on the meaning of clauses and generate recommendations.
The structured workflows inherent in these frameworks help reduce human error and accelerate analysis. The combination of document parsing, reasoning over relationships, and multi-agent orchestration is ideal for large-scale, complex document intelligence tasks where both depth and breadth of research are required.
Audit and Compliance
Use case: Responding to compliance and audit requests, cross-referencing regulations, and adapting to new requirements
This also sounds like a great fit for LangChain/LangGraph and CrewAI, where rule-based validations are key. LangChain/LangGraph can encode regulatory rules and requirements into structured queries or knowledge graphs, allowing automated cross-referencing and validation.
CrewAI adds the ability to create rule-based agents to execute compliance checks, flag deviations, and adjust workflows as regulations evolve.
These frameworks are great because they combine rule enforcement with flexible reasoning, which is important for audit and compliance where accuracy, traceability, and adaptability to new rules are needed.
Military Logistics
Use case: Adjusting procurement and distribution based on rapidly changing external influences like weather and transportation
This use case is where CrewAI shines.
CrewAI can simulate entire logistics networks, assigning agents to analyze or model a different part of the supply chain (weather, transportation, inventory). Agents can understand and map dependencies and suggest adjustments in real time.
This kind of multi-agent reasoning and simulation capability is key for military logistics because decisions must account for uncertainty and dynamic changes.
Citizen Services
Use case: chatbots that guide citizens through complex processes
This is a great fit for OpenAI Agent Framework/AgentKit with MCP.
OpenAI Agent Framework/AgentKit is great because of the conversational interface, similar to ChatGPT’s conversational interface, but adds the ability to orchestrate multiple tools behind the scenes. MCP enables the chatbot to maintain context across multi-step interactions, query databases, and follow workflows consistently.
Together, they allow the system to use natural language to guide citizens through complex procedures without overwhelming them.
NiyamIT developed a flood-risk advisor based on this conversational-AI pattern, allowing citizens to navigate risk information with natural language. As agentic AI capabilities grow, these systems will evolve from basic Q&A into orchestrating end-to-end analytical workflows.
Classified Intelligence Analysis
Use case: Processing and analyzing classified information across multiple security domains while maintaining strict compliance and audit trails.
The best fit is Microsoft Agent Framework, Google ADK, or AWS Bedrock AgentCore
For agencies handling classified workloads, enterprise frameworks shine because compliance, security boundaries, and audit capabilities are built into the foundation. If you're already operating in Azure Government, AWS GovCloud, or Google Cloud for Government, these frameworks integrate with your existing infrastructure, using the security controls, encryption, and access management you've already implemented.
Conclusion
Choosing the right framework depends on mission context: where your data must live, how much autonomy your agents should have, and how much transparency you require.
The real leadership challenge in the federal space isn’t just picking tools, it’s about understanding how to align your technical stack with the goals you’re trying to achieve.
As agentic AI matures, managing the flexibility vs. convenience trade-off will define how effectively government organizations use the next generation of intelligent automation. Sometimes you need Ohtani's flexibility and power, sometimes you need Freeman's consistency and ease, sometimes you need Mookie Betts' adaptability. Or if you're the Dodgers, you can just sign all three and get the best of every world.
In the last post in this blog series, we’ll look at best practices for implementing AI agent frameworks in government, so you don’t end up with a Giancarlo Stanton.
Footnotes:
(1) I’d make a Scott Boras “agent” joke, but he’d demand a 5 year base + 5 option period contract, a cost-plus-award-fee structure that somehow reimburses all ODCs, a $70M signing bonus, and some extra royalties based on “mission intangible assets” just for using his name, image, and likeness. To be clear, this is more of an LPTA blog…
(2) Or it might generate more questions than NASA SEWP.
(3) Maybe not modernize the VA’s electronic health records. Even Ohtani couldn’t fix that.
(4) The Red Sox had Mookie Betts, who could integrate into any team and play anywhere at an elite level (as a bonus - he’ll integrate right into your bowling team too), and traded him to cut costs in 2020. Since then, Mookie has won three World Series rings*, and the Red Sox have crashed out of the playoffs twice (2021, 2025) with three brutal AL-East last-place finishes in between. Lesson: don't sacrifice the right framework for short-term budget relief. Sorry, Bill Simmons.
*Technically three, but everyone knows the 2020 Mickey Mouse World Series doesn't really count. Sorry, Dodgers fans. Actually…not sorry at all, Astros fan here.
(5) Freddie won't be throwing 90+ mph splitters or playing shortstop anytime soon.
(6) Of course, not every front office functions this way. Consider the Yankees at the trade deadline, where everyone agrees they need a starter or a bat… and then Brian Cashman still finds a way to trade for a middle reliever with a 4.70 ERA. Your AI agent architecture should avoid that kind of roster logic.
(7) …and hope they don’t implode like the 2025 Mets bullpen in August.
This post is written by Rami Zheman, Senior Program Manager, Technology Consulting at NiyamIT (and as you may have already guessed, a big sports fan). It is part 2 in a 3-part series on how agentic AI is making government decision making more efficient. You can read the first part here and the last part here