Why More Context Makes Your Agent Dumber and What to Do About It — Nupur Sharma, Qodo

TL;DR

• 1

  LLMs often ignore the middle of long context windows: Sharma says Qodo's benchmarking shows a "U-curve" effect where models attend to the initial prompt and the latest inputs, while Jira details, MCP data, and other material in the middle get purged or overlooked.

• 2

  Context engines help, but they are expensive to scale: She describes a context engine as a "bouncer" that ranks what matters in a messy codebase, but says indexing across 600 to 700 repositories slows down and becomes unpredictable.

• 3

  Iterative retrieval is a practical default for many teams: Instead of building a heavy context engine, Sharma recommends a lighter "library card" approach that indexes topics so agents can fetch deeper code only when needed, with lower developer effort and decent results.

• 4

  The orchestration paradox wastes tokens on planning instead of doing: Powerful models like Opus can get stuck researching the best method over and over, hopping tool to tool and burning API spend without actually solving the task.

• 5

  An 80/20 hybrid model can tame agent loops: Qodo gives high-reasoning models the open-ended 80 percent of discovery and planning, then hands the final 20 percent, such as validation, summarization, and critic checks, to more deterministic models with timeouts or counter limits.

• 6

  Specialist agents plus a judge agent beat one giant overloaded agent: In Qodo's PR review flow, a context collector routes tailored inputs to security, code-diff, and Jira-focused agents, then a judge agent checks whether their outputs actually fit together and are relevant for the developer.