Hubsessed Office Hours — AI-Native

One job per agent: why long jobs drift

Nico Lafakis · Tuesday, October 6, 2026

Your AI agent follows the first instructions, then gets less accurate on a long job. Why, and what fixes it? The real fix is prompt structure. Splitting the job into two agents will help too, and I can almost guarantee it. A long job fills the agent's context window (everything it can hold in mind at once). By the time it reaches the extra instructions, the call, the synthesis and the next set of instructions are already in there, and the initial instructions are way back. "Once that's off, it's off." For a call summary, I'd use two agents, one job each. The first takes the call, does the synthesis and outputs the values. The second picks up those values, checks that they align, asks the first agent if they don't, and only then pushes them into the fields. Then test: "you just have to run a ton of testing." The full walkthrough, 2:49. Replay: https://www.hubsessed.io/office-hours/ai-native/2026-10-05