Which layer changed?

Portal, Spotify's internal developer platform, runs two declarative agents: one takes bulk file reading, the other takes predictable code generation such as tests, configuration scaffolding and type stubs. Both use Google's Gemini 2.5 Flash as the worker model. A Claude Code plugin called Shunt intercepts large reads through PreToolUse hooks and sends them to the bulk reader; files longer than 350 lines are blocked by default, while targeted reads with offsets and limits still pass through.[1]

The measurement covers four scenarios in a Java monorepo, and mean savings from bulk reading came to about 90 percent. The gain comes from moving input and output work out of the frontier model's context; the model itself did not change. Another reading is available: the saving may reflect the file sizes in this repository, and a codebase of smaller files might show a lower figure.[1]

Where did the new cost move?

Delegation carries its own bill. Network latency typically adds 10 to 30 seconds, and Portal caps an invocation at 30 seconds. Delegated analysis does not reliably return the line-level detail Claude needs in order to edit. Spotify says the worker models are not meant for debugging, architectural decisions or safety-critical reasoning, which leaves the choice of mode with the developer as a new decision to get right.[1]

On 1 September I placed the constraint on how much agent-assisted work a team can run in the execution boundary rather than in model capability; routing the reads opens the cost side of that same boundary. Portal leaves the worker model configurable, so a team can change the model or the instructions without touching the routing plugin.[1], [2]

What would test this?

Spotify published both modes through AiKA, and their reuse on repositories outside Spotify would turn the routing layer from one team's setup into a portable component. The signal to watch is plain: by 31 December 2026, independent teams reporting measured token savings from the same two modes on codebases other than Spotify's Java monorepo.[1]