AI Think Tank — Help Guide

Everything you need to know to run a multi-agent discussion.

1. Getting Started

Choosing a Provider & Model

Before starting a discussion, click the "Settings" toggle at the top of the page to expand the settings panel.

1. Provider — Choose which AI service to use. Supported providers:
   • Anthropic — Claude Sonnet 4.5, Claude Haiku 4.5
   • OpenAI — GPT-4o, GPT-4o Mini, o3-mini
   • DeepSeek — DeepSeek V3, DeepSeek R1
   • Google Gemini — Gemini 2.0 Flash, Gemini 2.5 Pro
   • Groq — Llama 3.3 70B, Mixtral 8x7B
2. Model — Select a model from the chosen provider. The dropdown updates automatically when you switch providers.
3. API Key (required) — The key for your selected provider. Get one from the provider's console. The placeholder updates to show the expected key format.
4. Brave Search API Key (optional) — Enables agents to search the web for live data and images. Get a free key (2,000 queries/month) from brave.com/search/api.

Click "Save" to store settings in your browser. API keys are saved per provider, so you can switch between providers without re-entering keys. They are never stored on the server — they go directly to the respective APIs.

Starting a Discussion

1. Select which agents you want to participate by clicking their chips (all are selected by default).
2. Type a topic in the input field at the bottom (e.g. "Should AI be regulated?" or "Analyze Tesla's Q4 earnings").
3. Optionally attach files for context (PDF, Excel, Word, images, code, etc.).
4. Click Start or press Enter.

2. The Speaker Queue

The Speaker Queue controls which agents speak and in what order. When a discussion starts, all selected agents are added to the queue automatically.

Queue Controls

• Play All — Automatically runs every agent in the queue one after another. Click again to pause.
• Next — Runs just the next agent in line, then stops.
• New Round — Advances to a new discussion round and refills the queue with all selected agents.
• Stop — Pauses auto-play (does not end the session).
• Clear — Empties the queue entirely.
• Shuffle — Randomizes the speaker order (The Mediator, The Judge, and Sentiment Analyst stay pinned at the end).

Reordering & Editing the Queue

• Drag & drop queue items to reorder (grab the ☰ handle).
• Use the arrow buttons (↑ ↓) to move an agent up or down.
• Click × to remove an agent from the queue.
• Use the "Add agent" dropdown to insert any agent into the queue at any time — even agents that aren't in the current selection.

Mobile Users

On mobile, the queue is collapsed by default. Tap the "Speaker Queue" bar to expand or collapse it. The chat area fills the rest of the screen.

3. Interacting During a Discussion

Once a discussion is running, the input field changes to an interjection box. You can type a message and click Send (or press Enter / Ctrl+Enter) at any time — even while an agent is speaking.

Your interjections are added to the transcript and every agent that speaks after will see and address your message. Use this to:

• Steer the conversation in a different direction.
• Ask agents to focus on a specific sub-topic.
• Request specific output formats (e.g., "Give me a table comparing X and Y").
• Ask agents to search for images or data (e.g., "Find me a chart of GDP growth").

4. Meet the Agents

Each agent has a distinct personality, expertise, and discussion style. Click an agent chip to toggle their participation.

• 🔬 Dr. Nova
  Science & Technology — Data-driven, evidence-based, cites studies.
• 🏛️ Philosopher Phil
  Philosophy & Ethics — Socratic questions, moral dimensions.
• 📊 Biz
  Business Strategy — ROI-focused, pragmatic, bottom-line thinker.
• 🎨 Creatia
  Creativity & Arts — Metaphorical, unconventional perspectives.
• 😈 Devil's Advocate
  Critical Analysis — Challenges every assumption head-on.
• ⚖️ The Mediator
  Synthesis & Consensus — Finds common ground, summarizes. Speaks last.

• 🔥 Rex Risk
  High-Risk Strategy — Bold bets, asymmetric payoffs.
• 🚀 YOLO Max
  Speculative Trading — Momentum-driven, high-conviction plays.
• 📈 Samantha Street
  Equity Research — Valuation-focused fundamental analysis.
• 🌎 Dr. Macro
  Macroeconomics — Systems-oriented, big-picture thinker.
• ⚙️ Systems Sage
  Systems Design — Architecture and engineering patterns.
• 🧠 Dr. Bias
  Behavioral Psychology — Identifies cognitive biases.
• 🏦 Capital Steward
  Capital Allocation — Long-term, disciplined investing.
• 🛰️ Atlas
  Geopolitics — Power dynamics and global strategy.

These personas represent morally challenging perspectives. They are designed to stress-test arguments, reveal blind spots, and surface uncomfortable truths about human rationalization.

• 🛋️ The Complicit
  Status Quo Defense — Passive, norm-following, uses legal compliance as moral licensing.
• 🧮 The Operator
  Transactional Pragmatism — Cold, efficiency-driven, treats exchanges as consumer services.
• 🌑 The Cynic
  Existential Critique — Disillusioned intellectual who sees the ethical catastrophe but views action as futile.
• ♟️ The Sovereign
  Amoral Optimization — Game-theoretic rationality, maximizes utility with zero ethical weight.
• 🎲 The Defector
  Conditional Nihilism — Strategic pessimism with a personal exit tripwire.

5. Example Topics

Not sure what to discuss? Here are some ideas organized by category:

6. Getting the Most from Discussions

Agent Selection Tips

• Finance topics: Use Rex Risk, YOLO Max, Samantha Street, Capital Steward, Dr. Macro, and Dr. Bias for a full spectrum from speculation to discipline.
• Ethics debates: Use Philosopher Phil, Devil's Advocate, The Complicit, The Cynic, and The Mediator to explore all angles.
• Tech/science: Dr. Nova, Systems Sage, and Biz give you evidence + architecture + business viability.
• General brainstorming: Start with 4-5 diverse agents. Add more if you want deeper coverage.

Multi-Round Strategy

• Round 1: Let agents give their initial takes. Don't interject yet.
• Round 2: Interject to focus on the most interesting disagreements. Agents will build on Round 1.
• Round 3+: Push for concrete conclusions, recommendations, or action items.

Interjection Techniques

• Be specific: "Compare the GDP data from Dr. Macro with Biz's profitability analysis" beats "discuss more."
• Challenge agents: "Dr. Nova, where's the evidence for that claim?" triggers deeper responses.
• Request formats: "Give me a pros/cons table" or "Rank these options 1-5" for structured output.
• Add data: "The latest CPI came in at 3.2% — how does this change your analysis?" keeps things current.

7. Sentiment Analysis

The Sentiment Analyst is a specialized agent that tracks where each panelist stands on the core disagreement in your discussion. It speaks at the end of every round (after The Judge), providing both qualitative commentary and a structured sentiment chart.

How It Works

1. The Sentiment Analyst is queued last in every round — after The Judge.
2. When it speaks, it analyzes the positions of all agents who spoke that round.
3. It identifies two main opposing viewpoints (the "points of contention") and scores each agent on a scale from -1 (fully aligned with Viewpoint A) to +1 (fully aligned with Viewpoint B), with 0 being neutral.
4. After speaking in the chat, it updates the sentiment chart automatically.

Setting Points of Contention

Click the 📊 Sentiment button in the header to open the full-page Sentiment Analysis panel. At the top, you'll see two input fields where you can define the opposing viewpoints before or during the discussion.

• Custom viewpoints: Enter two opposing positions (e.g., "Regulate AI strictly" vs "Let markets self-regulate"). These will be fixed for all rounds.
• Auto-detect: Leave the fields blank and the Sentiment Analyst will automatically identify the key points of contention from Round 1. Once detected, they are locked for all subsequent rounds and the input fields are auto-filled.

The Sentiment Panel

The panel (opened via the header button) is a full-page overlay containing:

• Viewpoint Configuration — The two input fields for setting points of contention.
• Spectrum Strip — A horizontal bar showing each agent's current position as their emoji avatar, positioned from left (Viewpoint A) to right (Viewpoint B).
• Time-Series Chart — A line chart tracking how each agent's position moves across rounds. Each agent is represented by their emoji and a colored line.
• Qualitative Analysis — Written commentary from the Sentiment Analyst about trends, shifts, alliances, and outliers. Use the round selector to view analysis from any round.

Reading the Chart

• The Y-axis runs from -1 (Viewpoint A) to +1 (Viewpoint B). The center (0) is neutral.
• Each agent is identified by their emoji avatar and a colored line matching their theme color.
• Outliers — Agents with extreme positions (far from the group average) are highlighted with a glowing halo effect on the chart and a pulsing animation on the strip, making them easy to spot.
• The legend at the bottom shows which emoji and color belongs to which agent.

Removing the Sentiment Analyst

Unlike The Mediator and The Judge, the Sentiment Analyst is optional. You can deselect it from the agent chips or remove it from the queue. If removed, the sentiment chart will not update for that round.

8. Attaching Files

Click "Attach" in the bottom toolbar to upload reference materials. Agents will read and cite them during the discussion.

Supported file types:

• Documents: PDF, Word (.docx), plain text, Markdown, HTML
• Data: Excel (.xlsx), CSV, JSON
• Code: Python, JavaScript, TypeScript, and other code files
• Images: PNG, JPG, GIF, WebP (metadata is extracted)
• Video: MP4, MOV, AVI, MKV (metadata only)

9. Saving & Loading Discussions

Saving

Use the save dropdown and button in the bottom toolbar:

• Save as HTML — A styled, self-contained HTML file you can open in any browser. Also embeds the raw data so it can be reloaded later.
• Save as JSON — Raw structured data (topic, messages, rounds, agents). Good for programmatic use.

Loading a Previous Discussion

1. Click "Load" and select a previously saved .html or .json file.
2. The full transcript will be displayed and the topic field will be pre-filled.
3. Click Start to continue the discussion. All agents will have context of the prior conversation.

10. Web Search & Images

When a Brave Search API key is configured, agents can autonomously search the web during their responses. They use this to find current data, cite sources with links, and embed images.

You can also explicitly ask for searches by interjecting, e.g.:

• "Search for the latest unemployment figures."
• "Find me a chart comparing iPhone vs Android market share."
• "Look up the S&P 500 performance this year."

Image results are embedded directly in the chat. If an image fails to load, a fallback thumbnail or link is shown automatically.

11. Token Usage & Cost Tracking

A usage bar appears below the settings panel once a discussion starts. It shows cumulative token usage and estimated cost for the current session.

• Tokens in — The total input tokens sent to the API (your topic, transcript, file context).
• Tokens out — The total output tokens generated by agents.
• Estimated cost — Calculated using the selected model's pricing. This is an approximation and may differ slightly from your provider's invoice.

12. Tips & Best Practices

• Start small: Begin with 3-5 agents for focused discussions. Add more for broader debates.
• Use The Mediator last: They synthesize everyone's points — best placed at the end of a round.
• Multi-round discussions: Click "New Round" after each pass. Agents will build on prior rounds.
• Be specific in interjections: "Compare the GDP data from Dr. Macro with Biz's profitability analysis" gives better results than "discuss more."
• Upload context: Agents give much better analysis when they have data to work with (PDFs, spreadsheets, reports).
• Use Devil's Advocate: They'll stress-test ideas the other agents are too agreeable on.
• Try the Shadow Panel: Add The Cynic or The Sovereign to ethics discussions for uncomfortably honest perspectives.

13. Keyboard Shortcuts

• Enter — Start discussion / Send interjection
• Ctrl+Enter — Submit from anywhere on the page
• Escape — Pause auto-play / Close settings panel
• Tab — Navigate between agent chips and controls
• Enter or Space — Toggle focused agent chip

14. Troubleshooting

• "Please add your API key" — Open the Settings panel, select your provider, and enter a valid API key.
• "Connection error" — The backend server may not be running. Ensure python main.py is active on port 8000.
• Search not working — Check that your Brave Search API key is valid and has remaining quota.
• Agent gives generic response — Try interjecting with specific instructions or upload reference documents for richer context.
• Images not loading — Some image sources block embedding. The app automatically tries fallback thumbnails. If all fail, a clickable link is shown instead.
• Session limit reached — Delete old chats from the History panel to make room for new ones.

AI Think Tank — Multi-agent discussion platform. Supports Anthropic, OpenAI, DeepSeek, Google Gemini, and Groq.