Capstone, three agents on a real problem
Run Research, Critic and Analyst in parallel on a topic from your life and synthesise the three reports into a recommendation.
Level 5 walked you through the multi-agent setup. You know what a CEO agent is, how workers run in parallel, where the neutrality guard kicks in, and how shared memory works without the agents poisoning each other. Now you do this once for real on an actual question from your life.
It's not theory and not a tutorial. You use three agents as decision support for something you actually have to decide.
The task
Pick a real decision you'll make in the next 30 days. Examples:
- Which job change to make if you're considering some.
- Should you build a premium feature for your side project or push for reach.
- Which of two trainings has the better ROI for you.
- Should you buy your apartment or keep renting.
- Which tech stack fits the small SaaS you want to build.
Important: it has to be a real decision, not a dry run. You'll have to use the synthesis at the end, otherwise you don't learn anything.
Setup
You need the three agents from Level 5: Research, Critic, Analyst. You can build them yourself from the patterns of these lessons, or you take the ready-made implementations from the nex-hq repo as a template. What matters is each agent has a clear system prompt and all three use the same MCP stack.
If you don't have your own MCP servers fitting your topic: take web research (mcp-research, brave-search, exa) and memory (StudioMeyer Memory or Mem0). That covers 80 percent of decisions.
What the agents should do
Research takes the question and delivers data, sources, comparison values. What other people say, what studies say, what the numbers are.
Critic takes the same question and actively looks for reasons why the answer you seem to want would be wrong. This is the most important agent. Its job isn't to confirm your idea, but to take it apart. If it says "all good", you formulated the critic prompt too weakly.
Analyst takes the question plus historical patterns from your memory. Its job is to draw comparisons with your past. If for example you've stood at exactly this decision twice before and decided wrong both times, the Analyst has to make that visible.
The output
Three reports as Markdown files (research/YYYY-MM-DD-{agent}-{topic}.md).
A synthesis you write yourself (not the CEO agent). 400 to 600 words. The synthesis must contain:
- What all three reports see the same.
- Where Research and Critic contradict, and your assessment why.
- What the Analyst pulls from your past and whether you consider it relevant.
- Your decision with three concrete next steps.
Plus a reflection (200 words): did the setup make the decision easier? Where was it help, where was it theater?
Self-grading criteria
- Does your critic report sound truly critical or did it let itself be intimidated by you? If critic mostly agrees, the system prompt is too friendly.
- Did the analyst use real patterns from your memory, or just generic tips? If the latter: memory is empty or you filtered the wrong tags.
- Is your synthesis more than the sum of the reports? A real synthesis has its own thought that none of the three agents would have alone.
- Did you actually decide? If "still don't know" sits at the end, the setup didn't work.
Why this capstone in particular
Multi-agent setups are often sold as a playful demo. Three agents debate something and at the end produce a nice report. That's show. The real value only shows up when you use it as a tool for your own decisions, because then you learn where the setup helps and where it's theater.
You'll find Research and Analyst are usually quickly helpful. Critic is the agent that disappoints most often because it answers too politely. When you find a critic prompt that actually makes you reconsider, you've won.
What comes next
In Level 6 you build your own MCP servers instead of just using others. With that, you can plug the agents in your multi-agent setup into your concrete data sources, instead of working with the generic tools. Three agents on web search become three agents on your customer DB, your notes, your codebase. That's exactly when the system gets truly personalized.