Field note
Spotify's Portal example treats coding-agent cost as a workload-routing problem rather than a single-model choice. Portal AiKA modes are declarative agents that run on ephemeral infrastructure with configured instructions, models, parameters, and MCP tools. The published shunt plugin uses pre-tool hooks to divert broad large-file reads and predictable code generation, while leaving debugging, architecture, and safety-critical work on the frontier-model path.
The mechanism combines an enforceable routing hook with reusable worker modes and concise return formats, reducing how much bulk input and output the primary agent consumes. This matters now because token spend is increasingly dominated by file I/O and routine generation rather than by the small fraction of steps that need stronger reasoning.
Teams can copy the pattern only after measuring threshold size, network latency, summary fidelity, and the cost of rereading exact sections before edits. The reported roughly 90 percent mean bulk-read saving came from one Java-monorepo test set; it is evidence for a routing pattern, not a universal benchmark. Watch whether agent platforms expose auditable routing decisions and task-level cost without relying on advisory repository instructions alone.