LinkedIn engineer traces agentic harnesses from chains to code

Most agent builds you read about share a shape. A prompt feeds another prompt. A graph routes calls between tools. A sandbox runs the agent's own code. In an InfoQ article this week, a LinkedIn engineer sorts those shapes into three stages. The piece names when each one is the right choice.

The piece is part of InfoQ's Agentic AI Architecture series. It uses LinkedIn's own agent builds as examples. Each example is tied back to a published LinkedIn engineering post.

The three stages InfoQ names

Chains come first. The author describes them as a line of prompts where the output of one step feeds the next. The example is LinkedIn's early RAG agent. The three steps were routing, retrieval, and generation. The piece says this shape is fast to build and has the lowest cost per call. It also says one bad step upstream can break the whole flow.

Graphs come second. The author says nodes represent tasks and edges hold the control flow. The piece links to LinkedIn's own LangGraph-based build of Hiring Assistant, a tool for recruiters. The build added conditional execution, retries, state management, and schema checks. The piece says this shape works for complex tasks. The cost is more model calls and more orchestration.

Code is the third stage. The piece defines a code harness as four parts. A sandbox runs the code. Instrumentation logs every step. A validation harness tests the output against rules. A refinement loop retries until the output passes. The LinkedIn example is an internal tool. Staff ask questions in plain English. The agent writes SQL and Python, runs it against sample data, and checks the output against privacy and governance rules.

Why the author says code wins where it fits

The piece puts the argument in one line. A language model's output is probabilistic and must be validated. Code is deterministic. It either runs or it fails. The author says the pass or fail signal gives tight feedback loops. Those loops drive reliability up. The LinkedIn team says the code approach got the internal tool to market faster. They did not have to write a bespoke prompt or a workflow graph.

The four rules the piece lands on

The author ends on four rules a team can apply at any stage. Treat every model output as untrusted input. Decouple planning from execution. Invest in observability. Build for continuous improvement with real feedback loops. The piece closes on a line a team buying agent work can hold the vendor to. Pick the least complex abstraction that meets the reliability, speed, and cost bar.

Source

The Evolution of Agentic Harnesses: From Chains to Graph to Code, InfoQ.