Agent engineering is proposed as a new discipline for building reliable LLM-based agents in production. Unlike traditional software, agents are non-deterministic and can behave unpredictably across multi-step workflows. The discipline combines product thinking (prompt writing, defining scope), engineering (tools, runtimes, UI), and data science (evals, monitoring, error analysis). The recommended development cycle is: build, test against imagined scenarios, ship early, observe production traces, refine prompts and tools, and repeat. Key challenges include every input being an edge case, inability to debug with traditional methods, and the fact that uptime alone doesn't indicate correctness. Companies like Clay, LinkedIn, Vanta, and Cloudflare are cited as examples of teams successfully applying this approach.
