Agentic Focus Group

A multi-agent system that simulates qualitative focus groups using GitHub Copilot custom agents.

Status: Exploration / Proof of concept — this is an experiment, not a production tool. See Limitations below.

What is this?

This repo defines a set of GitHub Copilot custom agents and skills that orchestrate a fully simulated focus group — from research brief to final report — with every participant impersonated by an AI agent.

You provide a research topic or question. The system produces a complete deliverable set:

workspace/YYYY-MM-DD-{slug}/
  brief.md                  # Research brief
  discussion-guide.md       # Moderator's topic guide
  participants/
    {name}.md               # One persona file per participant
  transcript.md             # Full moderated session transcript
  analysis.md               # Thematic coding & analysis
  report.md                 # Final report with recommendations

The methodology is grounded in real focus group best practices (see knowledge/focus-groups/).

Agents

Agent	Role
Lead Researcher	Orchestrator — plans the study, creates the brief, designs personas, builds the discussion guide, then delegates session and analysis to sub-agents.
Moderator	Facilitates the simulated session round by round, calling on each participant agent in turn.
Participant	Impersonates a single persona — responds in-character to questions and to other participants.
Analyst	Performs thematic analysis on the transcript and writes the final report.

Skills

Skill	Used by
create-brief	Lead Researcher
design-participants	Lead Researcher
create-discussion-guide	Lead Researcher
analyze-transcript	Analyst
write-report	Analyst

How to use

1. Open this repo in VS Code with GitHub Copilot enabled.
2. In Copilot Chat, select the lead-researcher agent.
3. The agent will execute all phases autonomously and produce the output folder under workspace/.

Sample Input

A beverage company is preparing to launch a new premium sparkling drink positioned between flavored sparkling water and light mocktails. Before finalizing the product and brand direction, they want to understand:
- How consumers perceive the emerging "adult non‑alcoholic" category
- What emotional and functional needs drive purchase
- Which brand narratives resonate (e.g., sophistication, wellness, indulgence)
- How packaging, naming, and flavor cues shape expectations
- What barriers or misconceptions might limit adoption

See the resulting report.

Orchestration flow

User → Lead Researcher
         ├── Phase 1: Create brief                    (create-brief skill)
         ├── Phase 2: Design participant personas      (design-participants skill)
         ├── Phase 3: Build discussion guide            (create-discussion-guide skill)
         ├── Phase 4: Run session                       (subagent → Moderator → subagent × N → Participant)
         ├── Phase 5: Analyze & report                  (subagent → Analyst)
         └── Phase 6: Quality check

Limitations

This is an early exploration. There are significant open questions and known limitations:

Subagent uncertainty

The architecture instructs the Moderator to "run a subagent @participant" for each persona in each round. However, it is unclear whether GitHub Copilot actually spawns a separate sub-agent per participant call, or whether the Moderator agent ends up simulating all participants within its own context. If the latter, persona separation is weaker — all "participants" share the same context window, weights, and potential biases, reducing the independence of viewpoints.

Other limitations

• Not real qualitative research. Simulated personas cannot replace real human participants. Insights should be treated as hypotheses, not findings.
• No statistical generalizability. As with real focus groups, but doubly so here — the "participants" don't represent a real population.
• Groupthink risk is amplified. A single underlying model tends toward coherence. Even with diverse persona instructions, responses may converge more than real humans would.
• Context window pressure. Long transcripts may exceed context limits, causing later rounds to lose awareness of earlier discussion. Quality may degrade in longer sessions.
• Persona fidelity. There's no guarantee the model consistently maintains distinct communication styles, biases, and viewpoints across many rounds.
• No validation loop. The system doesn't verify whether the "participants" actually behaved consistently with their persona files.
• Single session only. The architecture supports multiple sessions with different casts (for stratified populations), but this isn't implemented yet.

What it is good for

• Rapid ideation and stress-testing. Quickly explore how a topic might land with different audiences before investing in real research.
• Discussion guide prototyping. Generate and refine discussion guides that can later be used with real participants.
• Methodology exploration. Test multi-agent orchestration patterns in a structured qualitative research context.

License

See LICENSE.