Agentic Harnesses evolve from Chains to Graph to Code for reliable LLM systems

This method involved stringing together prompts sequentially, where the output of one step fed directly into the next.

While quick to implement, this simplicity inherently introduced fragility. A failure in any preceding step could cascade, derailing the entire process because there was no built-in mechanism for error detection or recovery, embodying a "best-effort" quality.

As needs grew more intricate, developers transitioned to graph-based harnesses. These introduced explicit structure using nodes for tasks and edges to manage control flow, allowing for conditional branching, parallelism, and retries.

This shift made control flow declarative, enabling reasoning loops where the LLM could navigate complex, long-running tasks. Graph systems also spurred the development of multi-agent setups, requiring protocols for asynchronous agent-to-agent communication.

A major development involved treating execution as a primary function via code harnesses. Instead of merely asking the model to "think through" a problem in natural language, these systems generate, execute, and refine code within a sandbox.

Code harnesses offer deterministic outputs, contrasting with the probabilistic nature of natural language. This pass/fail clarity provides the tight feedback loops necessary for significantly boosting system reliability.

When designing any agentic system, developers must treat LLM outputs as untrusted input and decouple planning from execution. Furthermore, it is vital to select the least complex abstraction that meets the necessary reliability, speed, and cost requirements.

For narrow, linear tasks, chains often remain the most cost-effective solution, providing the lowest latency by minimizing orchestration overhead. Graphs become suitable when conditional logic is essential, while code harnesses address workflows demanding verifiable outputs.