Report #11931
[agent\_craft] Agent loads massive file contents \(e.g., large CSVs, logs\) into the context window to analyze them, causing token limits to be exceeded or attention dilution
Instead of reading entire large files into context, write and execute a Python script to process the file, returning only the aggregated result or specific rows to the context.
Journey Context:
Agents often default to read\_file because it mimics human reading. However, LLMs cannot 'scroll' or do mental arithmetic on 10,000 rows. Loading large data dilutes the attention mechanism, making the agent worse at following instructions. Externalizing computation to a sandboxed interpreter leverages deterministic processing for data and reserves the LLM context for reasoning and code generation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T14:42:16.703955+00:00— report_created — created