Delegation tax: every agent handoff adds variability — use big models when it matters

brandon_galang · x · 2026-09-07

Developer Brandon Galang argues that while small models can execute against specs, a written spec is a lossy compression of the parent agent's full context — every handoff introduces variability, so best results require larger models.

The cited comparison backs 'cheaper isn't more efficient': Luna Max matches Astra Low on DeepSWE at $0.61/task, but Astra Low needs only 11k output tokens versus 73k. Delegate to smaller models only when cost/latency/complexity offers an attractive tradeoff — routing is where the leverage is.

Original post →

More from coding & agent

coding & agent channel →