When several AI agents work on a web task, each only sees information it is directly sent. Often one agent finds a fact another needs but never passes it along, so the answer is wrong.
A new approach runs the original workflow, then finds the missing piece. It adds one connection between the agent with the fact and the one that needs it, reruns affected agents, and decides whether to keep the edited answer by checking internal changes, not just the final answer.
On five test suites, it improved web tool use by 8.0% and reasoning and planning by an average of 5.7% over the fixed workflow, at about one-third higher cost. This makes agent teams more reliable without retraining.