Agent Framework, Agent Runtime, Agent Harness. They sound interchangeable but they’re not.

Agent Framework

When using any agent framework, you get a standard way of calling various LLM providers with agent abstractions. Think of this as the “standard interface” for interacting with LLM provider.

Frameworks give us pre-built patterns for the most common agent architectures. Want to implement a ReAct style agent? The framework has that built in. Additionally, we get middleware, structured output, chains, and tool calling out of the box. The main value they provide is abstraction. But if the abstraction is designed poorly, it will significantly degrade the performance of agents built on top of that framework.

Examples: LangChain, OpenAI Agents SDK, Google ADK, LlamaIndex, CrewAI

Frameworks are amazing for getting started, but they can become a pain as your agent gets more complex.

Agent Runtime

Runtimes are lower level than frameworks. While frameworks give you abstractions for building agents, runtimes handle the infrastructure concerns, how your agent actually executes reliably in production.

The main feature that defines a runtime is durable execution. Durable execution means your workflow saves its progress at key points, so it can pause and resume exactly where it left off, even after crashes, failures, or significant delays. This is like checkpoint systems. Your agent crashes at step 8 of a 10-step workflow? It resumes from step 8, not step 1.

Beyond durable execution, runtimes also provide streaming (real-time progress updates), human-in-the-loop support (pause for approvals), state management (persist everything across sessions), and the ability to handle long-running tasks that span hours or days.

Examples: LangGraph

Runtimes can power frameworks. LangChain 1.0 is built on top of LangGraph to leverage its runtime capabilities. The framework handles abstractions, the runtime handles reliable execution.

Agent Harness

A harness is a fairly new term. A harness is higher level than framework. It comes with default prompts, built-in tools, planning capabilities, file system access, context management, and more. A opinionated agent system that comes with batteries included.

Many of you used create-react-app. Same analogy can be applied with harness. With create-react-app we get a working frontend project up and running, without touching anything we can at least visit the webpage. Agent harness is kind of same thing it comes with all the standard tooling and capabilities. These built-ins save you days to weeks of work, and more importantly, team is now laser focused on your problem.

Claude Agent SDK is the first in this space. With just 5 lines of code a claude code agent is up and running and this significantly decrease the TTFF(Time to First Feedback). Then we just have to customise the harness with our requirements.

Agent space is still crowded. so many ways are being used to build agents and there are very few reliable agents out there. I think we need find out the MVC moment in agent architecture then we can see more and more reliable agents. But my prediction is Agent Harness will be the next big thing that will give a structure.

If you are building agents lets talk.