Production AI agents are increasingly surrounded by conventional software controls rather than relying on a sequence of prompts alone. An InfoQ architecture review divides that evolution into three patterns: linear chains, workflow graphs and code-executing harnesses.
Chains are quick and inexpensive for narrow tasks, but one malformed intermediate answer can corrupt every later step. Graphs make branches, retries, parallel work and state explicit. LinkedIn, for example, used graph-based orchestration and persistent messaging to support long-running recruiting tasks while adding schema checks, business rules and human approvals.
Code harnesses go further by letting a model generate and run code inside an isolated sandbox. Logs, tests and assertions provide direct feedback, and the agent can revise its approach after a failure. LinkedIn has used this pattern for an internal analytics tool that translates natural-language questions into SQL and Python while retaining access, privacy and governance checks.
Executable code does not make a model’s planning trustworthy by itself. Sandboxing limits damage, while tests verify only the properties they actually cover. The review recommends separating planning from execution, treating every model output as untrusted input and instrumenting each step. Its most useful design constraint is also the least fashionable: use a chain when a chain is sufficient, a graph when control flow must be explicit, and a code harness only when verifiable execution justifies the extra operational cost.