Seeing @SentientAGI's release of the SERA-Crypto technical report, it feels like the open-source architecture is redefining the performance boundaries of AI Agents.

In the DMind benchmark test, it outperforms Perplexity Finance and Gemini, only trailing behind Claude 4.5 and GPT-5 by less than 3%.

In the internal encryption analysis benchmark, it ranks first, surpassing all existing AI systems—GPT-5 Medium Reasoning, Grok 4, and Perplexity Finance.

—————————————————————————

Traditional ReAct systems with complex reasoning loops lead to excessive latency, inconsistent tool calls, and inability to run multiple API calls in parallel.

For example, when a user asks, "To what extent is Lido exposed to stETH de-pegging risk over the next six months?" The system needs to coordinate over 50 endpoints, including TVL trackers, staking APIs, on-chain traffic providers, derivatives data, and more.

SERA uses embedded matching for routing inference.

—————————————————————————

The SERA architecture rephrases each input query, embeds it, and then compares it with two separate embedding indexes: the tool index and the prompt template index.

The tool index contains descriptions of over 50 endpoints: market data APIs, TVL trackers, on-chain traffic providers, derivatives data, social sentiment APIs, and more.

The prompt index includes brief descriptions and associated templates for 11 categories of crypto queries.

By shifting routing and prompt generation to the embedding layer, SERA avoids ReAct loops, consistently selects the same tools for different query categories, executes multiple tool calls in parallel, and maintains an average latency below 45 seconds.

—————————————————————————

The fully open-source tech stack proves that architecture can surpass computing power.

When @SentientAGI's open-source system outperforms closed-source alternatives in cost and latency, the form of AI Agents will undergo a qualitative change.