Why More Context Makes Your Agent Dumber and What to Do About It — Nupur Sharma, Qodo
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A talk by Nupur Sharma from Qodo covering practical failure modes in LLM-based agents and how to address them. Key insights include the 'lost in the middle' problem where LLMs focus on the start and end of context but ignore the middle, the orchestration paradox where capable models loop endlessly researching methods instead of solving problems, and task overload causing agents to drop objectives. Solutions discussed include context engines with ranking logic, hierarchical summarization, knowledge graphs, iterative retrieval, self-correction with critic nodes, and an 80/20 hybrid approach pairing high-reasoning models for open-ended research with deterministic models for final validation. The talk also covers Qodo's multi-agent PR review architecture using specialized sub-agents coordinated by a judge agent, with context calibrated via PR history and compliance guidelines.