Agent Architecture and Context Management - AI industry research indicates that excessive knowledge and memory in language models create attention competition with primary tasks [1] - Standard ReAct loops continuously accumulate prompt histories, retaining failed steps and creating interference with original objectives [1][2] - Plan-and-Act frameworks address long-running agent context management challenges that exceed prompt-tuning capabilities by decoupling planning from execution [3] Execution and Performance Metrics - On WebArena-Lite benchmarks, a standard ReAct-style executor without a planner achieved a success rate of 36.97% [6] - Naively finetuned planners degraded performance, dropping the executor success rate to 20.60% due to mismatched site instructions [6] - Properly trained planners improved task execution success rates to 43.63% [7] - Continuous replanning after each action successfully replaced failed attempts in the plan, further increasing the benchmark score to 53.94% [8]

Karpathy warned about this months ago:"If agents had less knowledge or less memory, maybe they would be better."His point was that everything already inside the model's input competes with the task for attention.A standard ReAct loop is built that way, since everything that goes into the prompt is never removed from it.For more context, the ReAct pattern runs one model in a single loop.The model generates a thought about what to do next, takes one action, reads the observation that comes back, appends all t ...