BPFlow AI / News

Gemini Decisions API, agent architect certification and AlphaGenome 2

Issue date:

This issue covers an API for choosing the next step in an AI workflow, a skills assessment for AI agent builders and a model for genetics research.

This issue may include important events published earlier.

News archive →
01

Google launches Gemini Decisions API

Google introduced Gemini Decisions API and the Gemini Decision 1 model. They evaluate predefined options rather than generate free-form answers. The API is available through the Gemini API in Google AI Studio.

Key facts

  • A single request can include text, images, audio, video and PDFs totaling up to 256,000 tokens, along with up to 1,000 independent questions.
  • The API supports yes/no checks, selection from 2–255 options and scoring against an ordered rubric. Answers include a confidence score.
  • Google reports end-to-end latency of 150–250 ms for contexts up to 16,000 tokens. Pricing is $0.10 per million input tokens, with no output-token charge. A free API tier is available for prototyping.

An objective look at what matters

The API could simplify selection of the next step in an AI workflow when the options are known in advance.

Business processes

  • Request routing and tool selection — A system could direct requests to a suitable model or tool using predefined criteria.

Automation opportunities

  • Checking incoming materials: Teams could evaluate several conditions against a shared set of materials and send ambiguous cases to a person.
    Conditions: Options and criteria must be defined in advance.; Confidence thresholds and decision quality should be tested on the organization’s own data.

Impact on manual work

Some routine checks could be delegated to the system, but a reduction in manual work in a specific process has not been demonstrated.

Limitations and risks

  • The stated latency applies to Google’s specified conditions; other scenarios require measurement.
  • An inaccurate confidence estimate could lead to incorrect routing.

Sources

  1. Google Blog: Gemini Decisions API: Fast, calibrated multimodal decisions · Source date:
02

Google Cloud opens Professional Agentic Architect certification

Google Cloud announced that its certification is available to professionals who design, build and deploy AI agents on Google Cloud.

Key facts

  • The certification combines a proctored multiple-choice exam with hands-on Google Skills labs.
  • Preparation can begin at no cost: Gemini Agent Ready participants receive 35 learning credits per month for access to labs and courses.

An objective look at what matters

The certification gives teams another way to assess practitioners’ knowledge and hands-on skills before working on agents for multistep workflows.

Business processes

  • Training AI agent practitioners — Employers could use the exam and practical labs as one indicator of a professional’s preparation.

Automation opportunities

  • Building agents for multistep, multi-application workflows: Trained practitioners could apply their skills to developing such solutions.
    Conditions: The certification is available; applying the skills depends on the project and infrastructure.

Impact on manual work

The certification itself does not automate work, and outcomes from future projects have not been demonstrated.

Limitations and risks

  • A certificate does not guarantee the quality or security of a particular implementation.

Sources

  1. Google Blog: Get certified as a Google Cloud Professional Agentic Architect. · Source date:
03

AlphaGenome 2 predicts how DNA variants affect cell types

Google DeepMind introduced AlphaGenome 2 to predict how DNA sequences and their variants affect specific cell types and states. The model is available via API for non-commercial research, with a private preview for commercial researchers.

Key facts

  • The model was trained on human and mouse data, including more than 107 million individual cell profiles.
  • Google DeepMind reports that on the TenK10K dataset, the model identified 30% of causal variants, versus 16% for the original AlphaGenome fine-tuned on single-cell data. For identifying the direction of gene effects, the figures were 43% and 23%.

An objective look at what matters

Cell-specific predictions could help researchers select DNA variants and experiments for further investigation.

Business processes

  • Planning genetics research — Research teams could rank candidate DNA variants before conducting experiments.

Automation opportunities

  • Analyzing DNA variants: The API could generate predictions to help select promising variants.
    Conditions: Commercial research access is limited to a private preview.; Predictions require experimental validation.

Impact on manual work

The model could help with preliminary variant selection, but time savings for a particular team have not been demonstrated.

Limitations and risks

  • The model is intended for research only and has not been validated or approved for clinical or diagnostic use.
  • The comparative results come from a specific dataset and do not guarantee the same accuracy in other settings.

Sources

  1. Google DeepMind News: AlphaGenome 2: Deciphering how cells read DNA to better understand disease · Source date: