Death Loops in Research Agents
Product manager at Cohere on enterprise AI search infrastructure and deep research agents
Death loops are the core reliability tax on long running agents, because the same extra reasoning that makes a report richer can also trap the model in repeated planning, browsing, or self checking instead of finishing the task. The issue shows up mainly in the interface layer for products like North and Manus, not in a simple API call and response flow, which means the hard problem is managing multi step orchestration over 10 to 15 minute jobs, not just generating one answer.
-
In practice, this is what happens when an agent keeps reconsidering the plan, reissuing searches, or revisiting the same subproblem rather than committing to an output. The interview places this in the minority of cases, but it is exactly the failure mode that appears once products move from chat to autonomous research workflows.
-
The contrast with basic search APIs is important. Tavily and Exa are closer to fetch and ground infrastructure, while Parallel is positioned around longer research tasks with asynchronous workflows and structured outputs. The more the product owns planning and reasoning across many steps, the more loop control becomes a product differentiator.
-
This is also why domain specific data matters. The Cohere interview argues that better dedicated knowledge bases would make these systems faster and better. Cleaner sources reduce the need for the model to keep searching the messy open web, which lowers the odds of the agent wandering or repeatedly second guessing itself.
Going forward, the winners in deep research will be the teams that turn reasoning from an open ended process into a bounded workflow. That means tighter tool routing, better stopping rules, and more domain specific sources, so agents spend less time thinking in circles and more time delivering usable work products.