Imagine this scenario (probably familiar to many of you): a brand new project lands at your company, and suddenly everything needs to be checked with various team members. How should we tackle this project? What technological requirements and which team members do we need? Does this still fit within our schedule? It's a proper round of consultations, and it doesn't always run smoothly. Surely there must be a way to streamline this process?

You can probably see where this is headed: this is exactly what we're going to explore in this tenth installment of The Generative AI Playground (we've reached double digits!). We assembled a fully agentic AI team on Slack, one that consists of personalized agents with their own expertise and tools, capable of executing projects from start to finish.

This team didn’t just steal the spotlight at the latest edition of The Bar. They gave us a clearer glimpse into the future of agentic AI in the workplace.

Meet the team

Without further ado, we're proud to present: our agentic dream team, modeled after five real Raccooners!

• Overseeing everything is project manager Hannah: this agent routes questions to the right specialist and keeps the trains running on time. This project manager can create full-fledged Word documents and fire off emails.
• Jens takes on the role of technical analyst. This agent can tap into Google Search for inspiration and ask follow-up questions when things get murky—this agent always goes the extra mile.
• Dries serves as the technical architect. This agent thinks in blueprints and building blocks, explicitly tasked with cutting through complexity, making final decisions, and proposing system architectures.
• Tom acts as our resource manager and friendly planning guardrail. This agent ensures nobody gets overbooked by keeping tabs on availability and workload.
• Aagje handles the visuals as our graphic designer. With access to an Image Generation tool, this agent delivers creative visuals for any project.

Each of these agents was clearly given a role, an objective, and the tools that were needed to finish their respective tasks. Their office? A Slack channel where their interactions unfolded in real-time. Fully traceable, as if watching real colleagues bounce ideas off each other.

The project

A simple message explaining the project put our agent team to work:

The crucial point here is that these agents have a mind of their own. Once they receive their assignment, they're completely free to choose their own actions (which we'll explore below). There's no rigid, pre-programmed workflow calling the shots here. Project manager Hannah jumps into action immediately: the agent whips up a to-do list in no time and loops in the right agent for the next task at hand. Hannah will always be first to respond to messages from other agents—complete with to-do list updates and a call-to-action for the next team member, just like any self-respecting project manager would.

Since some technical input is the first order of business for this project, technical analyst Jens gets the call. This agent maps out a comprehensive technical blueprint:

However, Jens does his work very thoroughly, so that's far from the end of it. This agent has been explicitly instructed to always ask follow-up questions and constantly dig for more information:

Next up is graphic designer Aagje's turn. This agent delivers a concept visual for the projected sales:

Now that the technical analysis and project visuals are buttoned up, resource manager Tom is brought in to outline the timeline, necessary resources, and potential risks. This agent also goes into considerable detail:

To wrap things up, technical architect Dries is enlisted to design an architecture for this project. In this diagram, the agent brings together all the technical components that have been discussed:

Hannah then bundles all the input from the other agents into a properly formatted Word document—complete with introduction, project description, technical analysis, architecture, visuals, and timeline. This agent also requests one final crucial approval from the user before forwarding the project proposal to the client.

With a combined zero coffee breaks, bathroom visits, and disagreements, this team managed to handle a complete project solely through Slack. Brainstorming, designing, scheduling… and even emailing: Hannah will immediately suggest forwarding the project to the client, and since she's got all the right tools at her disposal, she can actually do it for you. All in under ten minutes, the client receives a document spanning over five pages.

Creating the perfect twins

As you'd expect, these weren't your run-of-the-mill static chatbots: each agent was powered by Google's Gemini 2.5 Pro, running on Google's Agent Development Kit—a brand-new framework for building tool-wielding agents.

Each of these five agents was built from three essential components:

• An LLM - The agent's 'brain.' While this demo used Gemini 2.5 Pro, virtually any LLM could work here.
• Instructions - The agent's playbook. Through carefully crafted prompts, each agent was shaped to mirror the working style, expertise, and personality of their human counterpart. This handbook also provided broad workplace context: preferred frameworks, go-to programming languages, standard software, and so on.
• Tools - Software integrations like Google Search or an Image Generator that let agents move beyond mere conversation into actual action.

Worth noting: some agents occasionally ask for user input or approval. While human-in-the-loop involvement during this demo was fairly minimal, we could easily tweak the agents to ask for more input from the user when needed - allowing for more interactive and richer collaboration with the agentic team in future iterations.

Not just fast, but fun to watch

Admittedly, watching agents chat with each other in a Slack channel and craft project proposals is entertaining. But beneath the surface lies a glimpse of something much more significant.

• Fully traceable thinking: every reasoning step is visible. No black boxes here. The agents maintain constant access to the full conversation context.
• Scalable teamwork: need an extra analyst down the road? Just spin up another Jens. Want to experiment with a different LLM for Dries? No problem: the agents are modular. Each agent can be tweaked without confusing or disrupting the other team members, or messing up the workflow.
• True collaboration: the agents aren’t simply responding to input; they are genuinely collaborating, passing off tasks and context like real teammates.
• Natural communication: the entire system operates in natural language, making it easy to follow... even for non-tech folks.

And perhaps most exciting of all: this represents just one of countless possibilities that experiments like this unlock. Why shouldn't a media team deploy an AI content strategist who never misses a trend, or a marketing department occasionally call on an AI campaign manager? In our latest whitepaper, Cutting through the agentic AI noise, we dive even deeper into other possibilities and practical applications of agentic AI.

Keeping the agents in check

To get these agents to interact so smoothly with each other, they had to be fine-tuned with surgical precision. It's been clear for some time now that prompting is the backbone of AI applications, but working with agents like these really drives that reality home. Crystal-clear prompt structure with section headers, explicit do's and don'ts—all focused on achieving consistency. A sloppy prompt can trigger a vicious cycle of subpar output.

Then there’s the challenge of working with cutting-edge tech: new frameworks with sparse online documentation, and AI programming tools completely oblivious to their existence… Imagine, we could barely use AI to build this demo! Essentially, you’re flying blind, and testing everything yourself. It’s not necessarily a roadblock, but it’s definitely something to factor in.

Heading into the future

This Slack team isn't just a cool visualization of agentic AI's power: it also shows, in a familiar setting, what it would be like to work alongside agents in the workspace. The practical value becomes clear pretty quickly here. Agentic AI is no longer just a buzzword thrown around left and right: here, agents collaborate to tackle full-scale, multi-step projects independently. This represents yet another compelling example of agentic AI's transformative power. Unlike previous AI tools that required constant human oversight, these agents operate with remarkable autonomy while remaining fully accountable.

Agents aren't replacing people—they're extending what teams can achieve. They're fast, focused, and surprisingly entertaining to watch in action. Now if only Tom the resource manager could book us a real vacation...

‍