Why does my agent re-read the same file and waste tokens?

Your agent re-reads the same file because each step has no memory that it already read it. Without a shared, addressable layer, the model reloads the full file every time it needs a detail, and you pay for the whole file again. The fix is to give the agent a place to reference files by a stable address instead of re-ingesting their contents.

What does re-reading actually cost?

A single 2,000-token file re-read four times in one task is 8,000 tokens for one file the agent already had. Across a session with dozens of files, this is often the largest single line in your bill, larger than the actual reasoning.

Why is caching alone not enough?

Prompt caching can lower the price of a repeated read, but it has two limits. It still costs more per token than fresh context, and it invalidates the moment the file changes. An agent editing code or documents changes files constantly, so the cache keeps breaking exactly when you need it.

How does a shared data layer fix it?

In a shared data layer, every file has a stable address. The agent reads it once into the layer, then references that address on later steps instead of pasting the full contents back into the prompt. A second read of an unchanged file becomes a cheap lookup rather than a full re-ingest, and changes are tracked so the agent only re-reads what actually moved.

adlass provides this over MCP: your files live in shared Spaces, and the agent works against them by reference, so repeated access costs a fraction of a full read.