# The One Thing Every AI Agent Framework Agrees On - Mislav Ivanda, Daytona

## Executive summary

The talk argues that while major AI agent frameworks (e.g., LangGraph, CrewAI, OpenAI Agents SDK) fundamentally disagree on core concepts like what an 'agent' is or how workflows are defined, they all converge on a single architectural necessity: secure code execution via sandboxing. The speaker details three methods for running agent-generated code—allow lists, embedded runtimes, and separate VMs—and strongly recommends using a dedicated, isolated virtual machine (VM) to prevent a compromised agent from affecting the host application. The ultimate advice is to prioritize defining the execution boundary before selecting a specific framework.

## Key takeaways

- Convergence on Sandboxing: Despite differences in implementation (Python vs. TypeScript) and definition (LangGraph defines agents as graphs; CrewAI defines them as role-playing groups), all leading AI agent frameworks are moving toward dedicated sandbox services for code execution.
- The Recommended Isolation Boundary: The most secure approach is using a separate computer or VM accessed via API. This ensures that even if the agent misbehaves or is compromised, the sandbox cannot interfere with the main application logic.
- Architectural Priority: The choice of the execution boundary (local vs. external sandbox) is the most critical architectural decision, as it defines the entire system's security and development stack. Frameworks are merely 'opinions,' while the boundary is a 'commitment.'

## Technical details

- Agent Framework Discrepancies: Frameworks differ significantly: LangGraph defines agents as traversable graphs; CrewAI defines them as role-playing groups; and the OpenAI Agents SDK defines them as loops around an LLM model. They also differ in language (Python vs. TypeScript) and who controls agent behavior (explicitly defined workflows vs. self-directed goal achievement).
- Code Execution Vulnerabilities: 1. **Allow Lists/Approvals:** Vulnerable because the model can still open sockets or modify the file system if compromised. 2. **Embedded Runtimes (V8/QuickJS):** While providing network/file system isolation for the generated script, the actual tool calls (e.g., database query, email sending) run outside the runtime, creating a vulnerability path.
- Sandbox Implementation: The recommended approach is using a separate VM, interacting via API. This isolation prevents a compromised step from compromising the entire application. Frameworks like CrewAI and AG2 are adopting this by removing local code interpreters.

## Practical implications

- When designing an AI agent system, treat the execution environment (sandbox/VM) as the primary security boundary, rather than focusing on the agent framework itself.
- Architects must plan for external, dedicated sandbox services to ensure that local code execution failures or malicious actions cannot compromise the host application.
- The integration pattern for sandbox capabilities is standardized (create, run, delete), allowing for easier swapping of underlying sandbox providers.

## Topics

AI Agents, Sandboxing, System Architecture, Security Engineering, LLM Frameworks, Cloud Computing, Mislav Ivanda (Daytona), AIE CODE summit

Source: https://www.youtube.com/watch?v=WWIFiQi9U4w
