Once is not customary, Amazon Web Services finds himself distant in this agentics race.
In the Landerneau of AI, it is the word in fashion. The agents are everywhere, but how to define them? “It is a model of orchestrator language that will invoke services to go and draw targeted information, for example weather data. It can also invoke a tool to trigger an action. Hence the term agent. But this is not its most used function”, replies Yannick de Kercadio, CTO Google at Daveo, Entity of the ESN Magellan.
It was previously possible to code a program to pilot this process. With the generative AI, the program in question will be generated via the LLM by leaving a prompt. In our example, the prompt will typically ask the model to provide the weather according to GPS coordinates. The model will conclude that it will have to determine the area in question before going to glean information on a weather service available on the web. To carry out an action, it will also be able to analyze the parameters of the corresponding application API to access it.
AWS, Google and Microsoft have all developed NO Code tools to create agents on the fly. On this ground, Microsoft and Google are currently in the lead in the elbow (see the table below).
| Strengths | Copilot Studio and Azure Service Agent | Amazon bedrock ai | Agent designer and vertex ai agent builder |
|---|---|---|---|
| Integrated AI platform | X | ||
| Data orientation | X | X | |
| Developer orientation | X | X | |
| Richness of the third -party tool palette | X | ||
| Rich preinstructed agents | X | X |
On the Google side, it is Green Ai Agent Builder who is historically highlighted. One of its main strengths is its ability to automatically read APIs to support third -party sources and tools. The solution fits (like its competitors) to the offers of the Provider Cloud Ecosystem. At Google, this is particularly the case with Bigquery who has no equivalent on the market. “Vertex AI Agent Builder is also ahead in advance of the orchestration of the experience of a conversational agent,” points out Yannick from Kercadio. Within agentspace, which offers Microsoft, multiple agents on shelves, Google recently unveiled an agent designer. A second NO code interface which allows the Citizen Developers to develop agents. Thirty preconfigured tools and actions are accessible.
With agentspace, Google multiplies connectors. “This brick (currently in beta, editor’s note) will be usable as a centralized platform via which agents can be shared within organizations, “explains Cyrille Marechal, Lean Machine Learning Engineer at Devoteam.
Amazon and Microsoft court the developers
For its part, Amazon benefits from an important community of user and well -supplied documentation. Despite its no code orientation, its solution, called Amazon Bedrock AI, remains above all designed for developers. On this point, it offers SDKs to integrate personalized tools into agents. This ultimately makes it a solution that is not very suitable for non-technical users. “AWS clearly goes a notch further in terms of development, much more than Microsoft”, confirms Dimitri Cabaud, Lead Data & AI within the Microsoft entity at Devoteam. A point of view that Yannick de Kercadio does not share for whom the editor of Redmond, via GitHub Copilot, is precisely cut for developers. “As for the data scientists, it will turn much more towards Google and Vertex Ai,” notes the interested party.
“Amazon is historically positioned as an integration hub into the different LLM”
Cyrille Marechal adds: “Amazon is historically positioned as an integration hub in the different LLM.” This allows you to adapt finely to the treatment to be carried out. Thanks to Claude, AWS is able to define a duration of reasoning in the second, where Azure Openai is much less granular by offering three levels (Low, Medium and High). In total, AWS offers a wide variety of models that can adapt to different configurations. “We find the same logic on the Microsoft side. Within Azure Openai, GPT-4 remains better on the images. As for O1, it is more robust on text management”, Pondere Stephan Durey, Deputy CEO at Magellan Partners and CEO of Daveo. Ditto on the agent side with no less than 200 models identified. On this point, the race is tight.
Alongside Azure Ai Foundry, Microsoft offers Copilot Studio to create unanswered agents. An environment, which unlike GitHub Copilot, is within the reach of profiles without development competence. “Thanks to this solution, business users will be able to scaffold agents, connect data sources, and launch actions on applications or databases by enjoying hundreds of connectors provided via the No code Power Automate development solution. This will ultimately allow agents who will be able to carry out actions, including a large number of applications, such as Salesforce, SAP, Cabaud. Guillaume Gérard, Head of Genai South & Central Europe at Capgemini, adds: “Copilot Studio also benefits from third -party AI tools offered by Microsoft. It is integrated for example in its Voicebot services.”
Langchain competition
“Among the alternatives to the builders of hyperscalers agents are, for example, AI raised, or Uipath which combines the Robotic Process Automation to LLMS”, adds Guillaume Gérard.
According to a study that has just published the French consulting firm Ai Builder, the use of the no code solutions of the three hyperscalers are not profitable for the time being. For what ? “Because it is possible to achieve the equivalent or even better with open source frameworks in the forefront of which Langchain,” replies Pauline de Lavallade, Head of Ai within Builders Research, the AI Builders study cabinet. “We believe in the low code and no code approach to develop agents. The fact remains that publishers must give time to refine their approach in order to automate orchestration, the creation of the prompt react (Reason + Act, editor’s note) and the integration of tools. “Main defect pointed out: the articulation between the reasoning of the LLM and the action to be implemented which is judged by Ai Builder as rather badly described by the solutions of the hyperscalers.
