How Do You Build and Connect AI Agents? Notes from Google Agentic AI Night
How do you build AI agents and connect them so they can take on a task together? Start by defining each agent's role, the instructions it should follow, and the tools it can use. For several agents to work together, they also need a way to exchange requests and pass results on to the next step.
On October 13, I attended Google Agentic AI Night at Google's Seoul office. In this article, I'll share a campaign-production example presented at the event, looking at how agents for planning, copywriting, and image creation were put together, and how A2A connected a separate agent to the workflow. As we follow the process, I'll also consider what to look at when applying it to our own work.
AI agents for planning, copy, and images
The example is called Creative Pitch Generator. Starting with a campaign idea, agents responsible for planning, copywriting, and image creation produce their respective outputs, which are then brought together.
The demo used ideas such as an animal wearing a jetpack. First, the Creative Director develops the concept and its rationale. The Copywriter then writes the text for publication, while the Visual Director creates an image using the brand guidelines. In the final Assemble step, the text and image are combined into one output.
Dividing responsibilities among several agents like this is called a multi-agent approach. Thinking of them as a planner, a copywriter, and an image creator makes the idea easier to follow.
What I want to highlight is how each agent's output feeds into the next task. When we use AI tools separately, a person copies the plan into another tool, then collects the text and image and combines them. In this demo, those handoffs were built into the workflow.
For a real campaign, we could include the target audience and key message in the planning output so both the copy and image agents work from the same brief. We would still need to check that the text and image describe the same product and offer.
Sharing brand guidelines
Let's take a closer look at the Visual Director, the image-making agent. In the demo, it loaded the brand guidelines, used an image-generation tool, then saved and passed on the finished result.
A “skill” here can be understood as a set of instructions to consult when doing a particular job. The related official Google tutorial also organizes brand rules as a skill that the agent reads when needed.
In everyday work, the contents are familiar: brand colors, the feel of photographs or illustrations, product placement, and wording or elements to avoid. Just as a person consults a brand guide, we can give an AI a set of standards to use repeatedly.
This also makes it clearer what needs to change. If the brand colors change, update the guidelines. If the generation method changes, check that functionality. Review the finished image separately.
It helps to distinguish lasting standards from information that belongs only to the current campaign. The core brand colors can live in the guidelines, while this week's discount or the campaign's end date goes with the specific request. That reduces the chance of carrying conditions from an old campaign into a new one.
Connecting the image agent with A2A
In the demo, only the Visual Director was deployed as a separate service. The rest of the workflow called it when needed. Image creation was an independent capability that other parts of the process could connect to.
The technology used for this was A2A (Agent2Agent), a common protocol for independently running agents to exchange work requests, progress updates, and results. It lets them communicate what was requested, whether the work is finished, whether more information is needed, and what the result is.
If several campaigns use the same image-making capability, its brand guidelines and generation functionality could be managed in one place and called when needed. This approach may also be worth considering when teams use different systems or have different people responsible for operations.
There is no need to split every task into a separate service from the start, though. Remote requests introduce network delays, authentication, and connection failures to handle. More agents also mean keeping an eye on model usage charges and retry costs.
A single agent may be enough for a short piece of promotional copy. If the same team operates the whole process, dividing tasks within one application is another option. I would first ask how often the same capability will be reused, and whether there is a reason to manage it separately.
Three things to define for your own work
How might this apply to our own work? I suggest picking one familiar campaign and starting with the way information is handed from one task to another.
The first thing is the information supplied at the start. If the request is simply “Create promotional material for our new product,” the AI has to guess the product's features and audience. Provide the product description, target customers, publishing channel, and confirmed offer together. Also decide which missing details, such as the price or campaign dates, must prompt a question.
The second is the output passed to the next task. The planner might provide the key message, the copywriter the channel-specific text, and the image agent the file and its specifications. Check where that file is saved and whether the next person or agent can open it.
Revisions are part of real work, too. If the plan changes, which text and images need to be made again? Each output needs to remain linked to the request that produced it, so an image from an earlier plan does not get mixed with new copy.
The third is the criteria for deciding whether the result is ready to use. Text length and image dimensions can be checked automatically. Whether the product description is accurate, the result fits the brand, and the text and image communicate the same offer may require a person to review them. That is why an approval step belongs between finishing the creative work and publishing it externally.
In our own workflow, how much should AI complete, and what should a person check?
When information is missing or image generation fails, we also need to know where the process stopped and what needs to be retried. Costs should include retries and the time people spend making corrections.
The demo showed several AI agents dividing the work, from planning to copy and images, to create one combined result. I see it as a starting point for reviewing the repeated handoffs and revisions in our own work. Choose something you produce regularly and map out who receives what and how far each role takes it. That should make the parts you want to entrust to AI more concrete.
If you are thinking about how to introduce or use AI at work, get in touch with AIB's “On-site AI Advice Clinic.” When you request a consultation, we will visit you, listen to the challenges you face on the ground, and explore ways to address them together.
Talk to AIB about adopting and using AI
References
Event information: [Seoul] Google Agentic AI Night
Related official tutorial: Build an Agentic Pitch Generator: Orchestration, Evaluation, and Analytics with Google Cloud
Agent communication protocol: Core Concepts