Agentic Loop Simulator 2D: Plan-Act-Critique Pipeline

Modern LLM agents don't just answer a prompt once — they loop through planning, calling external tools, and critiquing their own output before a task is accepted, retrying failed attempts up to a budget. This 2D companion renders that loop as a flat, pannable lane diagram: task tokens spawn on the left, drift through Plan, Act and Critique lanes, and either exit as a success or are routed back for a retry (or dropped as a permanent failure once the retry budget runs out). Tool-call latency is modelled as an exponential service time, critique outcomes as a Bernoulli trial, and the eventual-success probability follows the standard geometric-retry formula — the same math that governs real production agent reliability, cost, and throughput trade-offs — with a live history strip tracking success rate and token burn over time.