Three Modes of Working with AI

Bjorn framed the shift in three stages. The old way was a simple loop: human does something, waits, does the next thing, waits again. Every step required manual effort and the gaps in between were dead time.

The AI-enhanced mode, where most people operate today, changes the loop to human-AI-human-AI. You write a prompt, get an answer, write another prompt. It is faster but still sequential. You remain the single thread that every task must pass through.

Agentic workflows break the sequential chain entirely. Instead of one conversation at a time, you fire off multiple agents and they work in parallel. Each agent handles a distinct task, and they only come back to you when something needs a decision or a quality check.

“The old way was human wait human wait. AI enhanced is human AI human AI. But agentic and workflows, you are firing and they work in parallel because one AI, two AI, twenty AIs, it doesn't matter.”
- Bjorn Vuylsteker

You Are Still the Bottleneck

Even in the agentic model, the human remains the constraint. Bjorn was candid about this. You still need to set direction, define quality standards, and review output. The difference is that you stop spending time on execution and start spending it on orchestration.

Bjorn described his own daily rhythm: he has become something closer to a QA tester than a builder. He moves through screens, identifies issues, drops them into a working process for agents to solve, and while those agents are solving he is already looking at something else. The loop tightens because the waiting disappears.

Bjorn's Parallel Workflow in Practice

On a typical morning, Bjorn has 15 agents running on a dedicated MacBook Mini. Some agents are doing SEO research. Others are scanning for product improvements. One is running GEO analysis. Another is handling competitive intelligence. He reviews their output in batches, flags what needs a second pass, and moves on. The throughput is not about working harder. It is about removing the idle time between tasks.

lightbulbThe Mindset Shift

Think in workflows, not execution. Your job is to define the process clearly enough that an agent can follow it without you holding its hand through every step.

MCP and the Tool Ecosystem

The Model Context Protocol (MCP) is the connective tissue that makes agentic workflows practical. If a tool has an MCP integration, it means it can talk to your AI. Bjorn emphasized how accessible this has become: you do not need developers to set it up. The protocol is standardized, and the setup process is straightforward.

Bjorn walked through a set of tools that represent different points on the agentic spectrum, each suited to different roles and comfort levels.

Open Claude: a developer-oriented tool that lets you build your own AI agent network from scratch. High flexibility, higher learning curve.
MacBook Mini agents: Bjorn runs roughly 15 agents on a dedicated machine doing research, improvements, SEO, and GEO analysis around the clock.
Cursor: started as a developer assistant but has grown into a full agentic environment. It connects to PostHog for analytics, Figma for designs, and Jira or ClickUp for ticketing.
Claude Co-work: built for product managers. It is visual, does not talk about code, but runs the same agent patterns on tools that matter to PMs like documentation, project management, and communication platforms.

infoWhat MCP Means for PMs

When you see that a tool has MCP support, it means you can connect it to your AI workflows without writing a single line of code. This is what makes agentic workflows accessible beyond the engineering team.

Skills: Reusable Process Definitions

A skill, in the agentic workflow context, is a Markdown file that defines the steps an agent should follow, the context it needs, and the boundaries of what it should and should not do. Think of it as a recipe card: explicit enough that someone who has never done the task before can follow it, flexible enough to handle variation.

Bjorn shared a concrete example from WeGroup. They manage backlogs that differ by country. An AI agent does not know which country a ticket belongs to by default. But inside the skill definition, Bjorn added a simple rule: if you do not know which country this is for, ask me. That single line turns a generic agent into one that respects the team's real-world workflow.

“A skill is just a Markdown file which defines the steps and what context the AI needs and what it needs to do and doesn't need to do.”
- Bjorn Vuylsteker

Skills also compound over time. Once you finish a process and feel like the result was solid, you can ask the agent to create a skill out of what just happened. The agent writes the skill file itself, capturing the steps, the context, and the guardrails. The next time you or anyone on your team faces the same task, the skill is already there.

lightbulbSkills Write Themselves

After completing a process that produced good results, ask your agent to turn it into a reusable skill. It will capture the steps, context, and constraints so the workflow can be repeated by anyone on your team without starting from scratch.

The Orchestrator Mindset

The central message of Bjorn's talk was a reframing of the product manager's role. In a world of agentic workflows, you are not the person who executes. You are the person who designs the system of execution. You define what good looks like, set up the agents, write the skills, connect the tools, and then let the system run.

This does not mean the PM becomes passive. Quite the opposite. The quality of the output depends entirely on the quality of the process definition. Vague instructions produce vague results. A well-defined skill with clear boundaries, explicit context requirements, and specific success criteria produces consistent, reliable output across runs.

From Executor to Orchestrator

Consider the difference between manually triaging 40 support tickets every morning versus writing a skill that defines your triage logic, priority rules, and escalation criteria, then letting an agent process the queue overnight. You still set the rules. You still review edge cases. But the repetitive cognitive work is gone, and your morning starts with a sorted queue instead of a raw inbox.

The tools are ready. MCP makes them connectable. Skills make them repeatable. The only thing left is for product leaders to start thinking in workflows instead of individual prompts. Stop prompting. Start orchestrating.