Where cheap-model delegation stops

Spotify’s Portal adds two “modes” around Claude Code: bulk-reader sends large files to Gemini 2.5 Flash and returns a summary, and code-writer generates boilerplate straight to disk without the expensive model seeing it. Reported result: a mean bulk-read saving of around 90%, measured on a Java monorepo across four scenarios.

The number is the least useful part, and worth reading carefully: it is a mean saving per bulk read, not a reduction in a session’s total tokens. A tool that routes 90% of the bytes of the reads it intercepts still spends whatever the reasoning costs, and the article gives no before/after session totals. The headline collapses a per-operation ratio into a total, which is the same shape as any other harness-engineering-summary counter that does not say what it counted.

The enforcement is a hook, not an instruction

What makes it work is that the agent is not asked to delegate. A PreToolUse hook blocks any read over a configurable threshold (default 350 lines) and redirects it to the worker model. That is agent-mistake-fix-harness applied to cost rather than to correctness: an instruction to “use the cheap model for big files” is advice the agent will drop under pressure, and a hook that refuses the read is not.

The two boundaries are the finding

Both were discovered by running it, and both are stated as flat prohibitions:

• You cannot delegate editing. A summary does not carry reliable line numbers, and an edit needs them. The cheap model can tell you what a file does; it cannot tell you where to cut.
• You cannot delegate reasoning. The worker found surface-level patterns and missed a subtle thread-safety bug — the class of defect that is invisible in a summary precisely because a summary is a compression of what the code says, not of how it behaves under interleaving.

Debugging, architectural decisions and safety-critical code are excluded by policy on top of that, which reads less like caution and more like the same two boundaries restated for cases where being wrong is expensive.

The floor nobody mentions in the headline

Each delegation costs 10–30 seconds, capped at 30. So the saving has a floor: below some file size the round trip costs more wall-clock than the tokens are worth, and the article says so — “counterproductive for small reads”. Any threshold like the 350-line default is a guess at where that crossover sits for one codebase, and it is the number a reader should expect to have to re-measure rather than inherit.

What it connects to here

needle-tiny-tool-model is the same boundary drawn from the other end: a 45M model that does only tool calls and structured extraction, with a calibrated confidence head deciding when to escalate. Portal decides by file size, which is a proxy; Needle decides by the model’s own estimate of whether it is out of its depth. The second is the better shape, and the harder thing to build.

token-efficient-web-requests attacks the same cost from a different direction — asking the source for less rather than compressing what came back. Content negotiation has no summariser in the loop, so it has neither the latency floor nor the lost-line-numbers problem. Where both apply, it is the cheaper one.