Gemini 3.5Google Bets on Agent Workflow (AI column)
Watched yesterday's. Gemini announcement 3.5 Google I/O. Google has officially published it. 19 May 2026 The big part here, I think, is not the next benchmark race, but how the company formulates the next layer of models. In short, Google introduced the Gemini family. 3.5 starting with 3.5 Flash. A model. 3.5 Pro, according to Google, is already in use internally and should go rollout next month. Interestingly, Flash no longer looks like the big model’s “fast little brother.” It is a model for frontier-level agents and coding.
This is a strong signal of where the market is moving. Until recently, the main conversation was about “how well the model answers complex questions.” Now there is more and more talk about whether the model can perform a long task autonomously, decompose it into steps, use tools, run subagents, maintain context and bring the work to the end.
According to Google, 3.5 Flash bypasses Gemini 3.1 Pro on a number of coding and agent benches: Terminal-Bench 2.1GDPval-AA, MCP Atlas, CharXiv Reasoning. Google also says output speeds up to 4x relative to other frontier models. It’s interesting to test these claims in your own scripts, but the benchmark expectations for the model are high.
If you look at the model itself, you can see that the bet is made on long-horizon tasks. In the Google post, there are many examples not about “answer the question”, but about “do the work”: codebase support, preparation of financial documents, analysis of unstructured assets, migration of legacy codebase, generation of interactive UIs, parallel approaches to UX problems. This is no longer a chat as an interface to the model, but a model as an executor within the workflow.
It is interesting how they describe antigravity. (This is their IDE, which is cool in front-line tasks.). In conjunction with the updated harness 3.5 Flash must run collaborative subagents and perform multi-step coding tasks under supervision. In essence, Google is moving the idea of supervised execution: the model does not just offer a piece of a solution, but becomes the orchestrator of a set of actions that a person or system controls.
Enterprise examples are also illustrative. Shopify uses subagents for long-term data analysis, Macquarie Bank pilots onboarding on documents 100+ pages, Salesforce integrates Flash into Agentforce, Ramp applies multimodal understanding to invoice recognition, Xero automates multi-step tax workflows, Databricks looks towards diagnosing problems in complex data environments. These are, of course, cases from Google material, but they show the direction well: models drag into the routine of companies with a lot of documents, data, checks and tool calling.
Another important detail: Google is not just bringing agentic layers into developer tools. 3.5 Flash becomes a default model in Gemini app and AI Mode in Search, and the new Gemini Spark is described as a personal AI agent that works 24/7 and acts on the instructions of the user. That is, the same logic goes in three directions: developers, enterprise and mass consumer products.
You should also be careful about safety. Google announces Frontier Safety Framework, enhanced cyber and CBRN (Chemical, Biological, Radiological, and Nuclear) Safeguards, more advanced safety training and interpretability tools. These are the right words, but the real security we will see when the model gets into the hands of millions of users, developers and others.
Anyway, Gemini. 3.5 It shows that the race of the top labs shifts to the agent model, where the frontier models are used to perform autonomous tasks, through decomposing them into pieces, using tools, launching subagents, supporting context and everything else ... in order to bring the work to a conclusion:)
#AI #Google #ML #Engineering #Software #Management #Agents