𝗔𝗴𝗲𝗻𝘁 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗜𝘀 𝗔 𝗖𝗼𝗺𝗽𝘂𝘁𝗲 𝗔𝗹𝗹𝗼𝗰𝗮𝘁𝗶𝗼𝗻 𝗣𝗿𝗼𝗯𝗹𝗲𝗺
Three independent groups recently arrived at the same conclusion for AI agent design.
Anthropic released a blog post on the advisor strategy. They use a cheap model to run the main loop. They only call an expensive model when the cheap one gets stuck. This setup at BrowseComp hit 41.2% accuracy at only 15% of the cost of using a top-tier model for everything.
Tobi Lutke from Shopify shared a similar setup on X. He runs a local model for research and uses a frontier model as an advisor. Developers built open-source versions of this within hours.
HazyResearch published a paper on a compressor-predictor framework. A small model distills context for a large model to reason over. Their system recovered 99% of accuracy at 26% of the cost.
This convergence is not a coincidence. It follows a specific design law: the cost-curve frame.
I have argued this frame across three layers in this series:
- Layer 1 (Retrieval): Why tool-loops beat RAG for most code tasks.
- Layer 2 (Storage): Why SQLite beats vector databases for symbol graphs.
- Layer 3 (Orchestration): Why the advisor strategy wins for model choice.
The logic is the same. Most agent tasks consist of many low-value operations and few high-value decisions.
If you use an expensive model for every token, you waste money on routine work like reading context or formatting text. The advisor strategy separates these paths. You use a cheap executor for the bulk work and an expensive advisor only for the critical decision points.
If you are building agents, watch out for these three engineering challenges:
- Data Egress: Sending context to a remote advisor can leak sensitive data. Use a redaction layer.
- Escalation Policy: Deciding when to call the advisor is hard. Too early wastes money. Too late wastes time.
- Handoff Design: The advisor should provide a short plan, not a full solution.
The pattern is real because it is efficient. Stop paying frontier-model rates for tokens that do not need them.
Optional learning community: https://t.me/GyaanSetuAi