Long tasks like web research require a large language model (LLM) agent to plan, reason, and check its own work. A single agent tends to verify too late, and its growing history buries what is currently relevant.
Goal-Aware Dynamic (GAD) gives each job an agent with its own context, keeps original goal and acceptance conditions visible, and uses a subtask-dependency graph to decide when agents act and what results they receive. checks subtask and final results, triggering repair when checks fail.
In experiments on long-horizon benchmarks, GAD improved success by up to 6.7 percentage points over the strongest baseline at lower cost than multi-agent baselines. This matters because LLM agents increasingly finish long tasks without human intervention.