InfoQ traces agent harness evolution from chains to graphs to code

InfoQ Chinese published an article titled “The Evolution of Agent Harnesses: From Chains, Graphs, to Code” on November 6, 2026, tracing a shift in the large language model agent space toward what it calls the agent Harness, also known as a structured execution framework. The piece argues that this evolution was driven by steadily growing demands from real business scenarios and unfolded in three stages: early systems were simple linear chains of prompts, intuitive and fast to ship but lacking control; then came tool-augmented approaches, followed by structured control. Its central claim is that the value of a Harness lies in the deterministic execution scaffolding it builds around a probabilistic model, pushing the model toward stability, observability, and production-grade quality.

To illustrate the fragility of the chain approach, the article draws on an early practice at LinkedIn: a generative AI agent built as a fixed three-step chain on LangChain. The flow began with routing, a classification step that determined whether a query was in scope and which specialized AI expert should handle it, then moved to retrieval, a recall-oriented step that called appropriate tools such as Bing web search or LinkedIn internal APIs, and closed with generation. Separating routing from retrieval reflected an early grasp of task decoupling, but the harness itself was little more than glue code for tool calls and conversation memory, treating the LLM as a black-box text function: text in, text out, plus a bit of syntactic sugar. This is all the reporting currently available on the topic.