Open Agent Specs

1. Multi-Agent Orchestration Pipeline

When you ask a crypto-related question, Assistant Hub routes it through a 3-agent pipeline. Agents A and B run in parallel for speed, then Agent C synthesizes their outputs into a structured report.

2. Agent A — Sentiment Analyst

You are a crypto market sentiment analyst.
Focus ONLY on: market narrative, social sentiment, community positioning,
fear/greed signals, news catalysts, and short-term trader psychology.
Be concise (3-5 sentences). End your response with:
Sentiment Signal: [Bullish/Bearish/Neutral] | Confidence: [High/Medium/Low]

Agent A specializes in reading the crypto market's "mood" — social media sentiment, fear/greed signals, news catalysts, and trader psychology. It outputs a directional signal (Bullish/Bearish/Neutral) with a confidence level.

3. Agent B — Technical Analyst

You are a quantitative crypto analyst.
Focus ONLY on: price action, technical indicators (RSI, moving averages),
volume trends, on-chain metrics, support/resistance levels, and chart patterns.
Be concise (3-5 sentences). End your response with:
Technical Signal: [Bullish/Bearish/Neutral] | Key Level: $[price or "N/A"]

Agent B focuses on quantitative analysis — price action, RSI, moving averages, volume trends, support/resistance levels, and on-chain metrics. It provides a technical signal with a key price level to watch.

4. Agent C — Synthesis

You are a senior crypto research synthesizer.
You receive sentiment analysis (Agent A) and technical analysis (Agent B) from two
specialized agents, then combine them into a final structured report.

You MUST respond with ONLY a JSON object in this exact format (no markdown, no code fences):
{
  "outlook": "Bullish" or "Bearish" or "Neutral",
  "conviction": "High" or "Medium" or "Low",
  "keyInsight": "1 sentence combining the most important signal from both agents",
  "actionWatch": "1 sentence on what to monitor in the next 24-48 hours",
  "riskNote": "1 sentence on the main risk to this thesis"
}

Agent C receives the outputs from both Agent A and Agent B, then combines them into a final structured report with an outlook, conviction level, key insight, action watch, and risk note.

5. Provider Router

The Provider Router manages connections to all LLM providers with built-in resilience:

• Circuit breaker — If a provider fails, it's marked unavailable for 2 minutes before retry. This prevents cascading failures from overwhelming a struggling API.
• Availability cache — Provider availability is cached for 30 seconds to reduce overhead on repeated health checks.
• In-flight deduplication — Concurrent requests to the same provider are deduplicated to prevent redundant API calls.
• Automatic fallback — When a primary provider is unavailable, the router automatically tries the next provider in the chain.

6. Intent Classifier

Before routing to a provider, each user message is classified into one of 7 intent categories. This enables smart routing — sending questions to the most appropriate model.

7. Autonomous Agent System

Users with Premium tier can deploy autonomous agents that monitor markets and execute paper trades based on predefined conditions.

Evaluation Loop

• Frequency — Every 60 seconds
• Startup delay — 15 seconds (lets price caches warm first)
• Data source — Shared market snapshot from cached prices, Fear & Greed, BTC dominance, and Pulse sentiment (zero additional API calls)

Condition Engine

Agents evaluate entry and exit conditions using a comparator system:

Safety Limits

• Maximum 5 agents per user
• Maximum position size: $100,000 (virtual)
• Minimum cooldown: 60 seconds between actions
• Entry conditions: ALL must be true (AND logic)
• Exit conditions: ANY can trigger (OR logic)

8. Persona System

Users can create custom AI personas with distinct communication styles and expertise areas. The persona system includes:

• Memory feedback loop — Personas learn from marked positive and negative response examples, using keyword-based retrieval to improve future responses.
• Context injection — Persona instructions, style guidelines, and relevant memories are injected into the system prompt for each conversation.
• Custom system prompts — Each persona has a user-defined system prompt that shapes its behavior and expertise focus.