Back to News
Advertisement

Concerned about how companies will manage their cloud bill once agents dominate

aabhijeet_gupte about 2 hours ago 1 comments

RU version is available. Content is displayed in original English for accuracy.

Every organization wants to become agentic. Corporations are laying off employees and diverting the spend to AI. Many are aware of the nuances of tokken maxing and how to manage their tokens.

My worry is what happens when agentic adoption normalizes and the underlying data layer starts generating huge bills for the compute as it is cloud based. Agents optimize for outcome. I can imagine a future where hundreds of thousands of agents operate on a data that shoots not only the token, but also the compute bill through the roof

Is there a holy grail of enterprise architecture that doesn't rip us off in the near future?

Advertisement

⚡ Community Insights

Discussion Sentiment

50% Positive

Analyzed from 120 words in the discussion.

Trending Topics

#agents#cost#agentic#deterministic#checker#database#model#agent#data#architectural

Discussion (1 Comments)Read Original on HackerNews

amit1525about 1 hour ago
The architectural solution that would keep the cost of agentic compute from growing exorbitantly is the migration towards MAS architecture that is based on a deterministic Maker-Checker loop. Rather than empowering giant agentic agents with unlimited capabilities to search and process cloud database information, companies opt for multiple smaller and specialized agents that use standard communication methods such as Model Context Protocol (MCP). In this model, a routing agent sorts requests, a worker agent performs them using the data that is available within a strictly defined scope, and a deterministic checker verifies the result prior to any database read/write operations. This way, there is no need for agents to fall into costly and endless "thought loops" and retry hallucinations, and the quantization of vectors in databases (for instance, 8-bit compression) reduces the cost of data layer processing significantly.