That "33k tokens before your prompt" study everyone shared

You probably saw the comparison: one coding agent harness sends ~33k tokens before your prompt, another sends ~7k. Big thread, lots of outrage about waste. I finally read the whole study instead of the headline, and the actually useful findings are different from what got shared.

First, in their realistic-config lane (instruction file + several MCP servers), the "light" harness came out HEAVIER: ~90.8k vs ~75k. A 72KB instruction file alone added ~20k tokens to every request, on both harnesses. Their own conclusion: configuration, not the harness, accounts for most of the production bill. The harness sets the floor, you set the ceiling.

Second, and this is the one that changed how I think about it: the cache behavior gap was way bigger than the size gap. The light harness kept its prefix byte-identical and wrote ~1,000 tokens to cache over a 5-request task. The heavy one kept rewriting its own prefix mid-session and wrote ~54,000, with single rewrites burning 43k+ at the premium write rate (cache writes cost 1.25-2x list depending on TTL, reads are ~10%). A stable big preamble is close to a fixed cost. An unstable small one can out-spend it. Size isn't the sin, churn is.

Third, session shape flips the winner anyway. On a multi-step task the heavy harness finished cheaper (121k vs 132k) because it batched tool calls. Rerun on a different model, it inverted (298k vs 133k). Subagent fan-out was a 4.2x multiplier. And their quality check found zero difference: both passed 5/5, one spending ~4x the tokens. So the honest answer to "which harness is cheaper" is "depends what your sessions look like", which is boring but true.

The part you can actually use: measuring your own takes two minutes. Most CLIs have a print mode with JSON output. Ask for something trivial, then sum three usage fields: uncached input + cache writes + cache reads. That's your preamble. I ran it on mine: 31,782 tokens in an empty directory, and my heavily configured project (MCP servers, plugins, a pile of skills) added exactly 166 more, because this harness version lazy-loads tool schemas. Config CAN dominate, and lazy loading CAN neutralize it. The probe tells you which world you're in.

Two caveats since numbers travel badly: it's a single-machine study with single-digit runs per lane, and my probe is n=1 on a different version. Portraits, not specs.

What do your numbers look like?