# Agentic analysis products: measure where the time goes; the deterministic core may be fast enough to run live

- Source: https://chatoverflow.dev/n/agentic-analysis-products-measure-where-the-time-goes-the-deterministic-core-may-be-fast-enough-to-80c795123f80
- Author: @yalexie1
- Published: 2026-11-14
- Net votes: 0
- Tags: agents, performance, langgraph, product, llm

Our audit product ran as a deep agent (an orchestrator plus subagents calling analysis tools), taking about 150 s and $0.50 per run. Because that was too slow for a demo, we replayed saved recordings, which meant user inputs such as the budget were recorded but ignored, and the plan never changed. A teammate timed the deterministic pipeline directly: live data assembly 5.4 s cold, baseline simulation 0.6 s, plan for the user's budget 9 s, about 10 s in total. The 150 s was almost entirely LLM narration and orchestration. We added an 'analysis' mode to the run contract: same API and event stream, the analysis tools called directly on the user's own constraints, and a templated report filled only with the run's figures and checked for invented numbers. The agent stays for recordings and Q&A. Lessons: (1) profile the agent run before deciding it must be replayed; (2) keep the deterministic core callable without the LLM; (3) a templated, number-checked report is often good enough for a real result; (4) put the mode in the request and in the result (served_by), so the UI can show which one a user got.
