Appendix - Agent Frameworks
What We Learned
- Frameworks package loops, tools, state, validation and observability; they do not replace system design.
- Choose abstraction level from branching, persistence, approval and deployment needs.
- Compare all seven frameworks on the same tool-using task and approval boundary.
- Distinguish delegated tools from handoffs, and conversational state from durable execution.
- Test with controlled models and validate approvals, recovery, traces and costs.
Key Acronyms, Concepts & Jargon
| Term | Short meaning |
|---|---|
| SDK / ADK | Software Development Kit / Agent Development Kit: tooling for building applications / agents. |
| LangChain / LangGraph | Model/tool integration layer / stateful graph orchestration. |
| OpenAI Agents SDK | Agent framework with tools, handoffs, guardrails and tracing. |
| Claude Agent SDK | Programmable access to an agent harness with tools and permission controls. |
| Google ADK | Agent toolkit with composition, runners and session management. |
| CrewAI crew / flow | Role/task-based team / explicit event-driven control flow. |
| Microsoft Agent Framework | Agent and workflow abstractions for orchestration. |
| PydanticAI | Typed agent framework with validated outputs and dependency injection. |
| Node / edge / state | Execution step / transition / data carried through a graph. |
| Checkpoint / session | Saved execution state / conversation-associated state. |
| Middleware / hook / callback | Extension points around model or agent execution. |
| Handoff / agent-as-tool | Transfers control / calls an agent and returns its result. |
| Deferred tool approval | Pauses a tool request until an explicit decision is supplied. |
| Dependency injection | Supplies tools with external services or configuration through a typed context. |