Agent Builder Reference

Everything you need to design, configure, and refine your autonomous trading agents.

---

​

What Is Gigabrain?

Gigabrain is an autonomous crypto trading agent platform. You create agents that:

1. Monitor markets using AI-powered specialist analysis
2. Fire on triggers (scheduled or news-driven)
3. Produce analysis and optionally execute trades on Hyperliquid

Each agent has: a goal, instructions, constraints, and triggers. The AI engine provides the analytical backbone — domain specialists that the agent can query in parallel to form a thesis.

---

​

Agent Structure

An agent is defined by:

​

Triggers

Agents support multiple triggers simultaneously: Scheduled triggers — run on a recurring basis:

• "every 4 hours" — checks markets every 4 hours
• "every 30 minutes" — high-frequency monitoring
• "every 1.5 hours" — fractional hours supported
• "daily at 9am" — once daily at a specific time

Alpha triggers — fire when matching market news/intel arrives:

• Uses semantic matching to compare incoming news against the agent’s focus area
• Configurable similarity_threshold (0.0–1.0): how closely the news must match
• Configurable min_confidence (0.0–1.0): minimum news quality to trigger on
• Lower threshold = more triggers, higher threshold = only highly relevant news

Manual triggers — you explicitly run the agent on demand.

​

Memory & Learning

Agents can be configured with memory enabled, allowing them to learn and improve over time:

• Memory stores learnings, not logs. The agent saves insights that should change its future decisions — patterns discovered from outcomes, not raw data or position history.
• Each run builds on previous runs. The agent sees its last 3 runs for recent context, plus can recall any relevant insight from its full memory.
• Insights are curated, not accumulated. The agent updates insights when its understanding evolves, deletes insights that are proven wrong or no longer apply, and periodically reviews its full memory to clean house.

What gets saved as a learning:

• “Tight stops on ETH get stopped out in high-volatility regimes — use wider stops when ATR is above X”
• “Funding extremes on SOL tend to resolve within 4-8 hours, faster than BTC”
• “Pre-FOMC squeezes on BTC tend to fade within 2 hours if no follow-through volume”

What does NOT get saved (available from other sources):

• Position entries/exits (in session history)
• Current positions/balances (fetched live from exchange each run)
• Market data like OI, RSI, funding (fetched live each run)

How it works in practice:

• Run 1: Agent spots a squeeze setup, takes a trade
• Run 2: Trade hit stop loss — agent checks why, notes “the structural level was too far from entry”
• Run 3: Similar setup appears — agent recalls the learning, requires tighter structural alignment before signaling
• Over time: The agent builds a personalized knowledge base of what works for its specific strategy

Enable memory for agents that:

• Trade repeatedly on the same setups (squeeze, regime, narrative)
• Need to adapt to changing market conditions
• Benefit from pattern recognition across many runs

Memory is less useful for:

• Simple monitoring/alerting agents that don’t make trade decisions
• One-off analysis agents

---

​

Available Specialists

Your agents can query these domain specialists:

​

Macro Specialist

• Domain: Macroeconomic regime, Fed policy, yields, DXY, risk environment
• Provides: Risk-on/off classification, upcoming catalysts (FOMC, CPI, NFP), gold/silver prices, macro indicators
• Best for: Regime-based strategies, pre-catalyst positioning, understanding macro backdrop
• Time relevance: Days to weeks

​

Positioning Specialist

• Domain: Perpetual futures positioning — open interest, funding rates, liquidation levels, whale flows
• Provides: Bullish/bearish positioning signals, crowding metrics, long/short ratios, taker flow, liquidation clusters
• Best for: Squeeze detection, positioning extreme reversion, understanding who is trapped
• Time relevance: Hours to days
• Supports timeframes: 1h, 4h, 1d, 7d

​

Fundamentals Specialist

• Domain: Protocol and chain health — TVL, revenue, fees, growth metrics
• Provides: Chain TVL trends, protocol revenue, sustainability metrics, sector positioning
• Best for: Mid-cap protocol analysis, fundamental screening, narrative validation
• Note: Understands protocol types — won’t mention TVL for AI protocols where it’s irrelevant

​

Technical / Price Structure Specialist

• Domain: Technical analysis and price structure — multi-timeframe
• Provides: Support/resistance levels, trend direction, momentum indicators, signal confluence, volatility
• Key feature: Multi-timeframe analysis with confluence and conflict detection across 15m/1h/4h/1d
• Best for: Entry/exit levels, invalidation prices, trend identification
• Time relevance: Minutes to weeks depending on timeframe

​

Market Regime Specialist

• Domain: Overall market regime and risk sentiment
• Provides: Risk-on/risk-off/neutral classification, macro drivers, aggregated crypto market structure
• Best for: Portfolio-level decisions, regime shifts, risk allocation

​

Social Intelligence Specialist

• Domain: Micro-cap social intelligence (tokens ranked 500+, market cap < $100M)
• Provides: Twitter/X sentiment, KOL activity, narrative identification, community health, red flags
• Best for: Meme coins, AI agent tokens, new launches, narrative momentum
• Note: For tokens without institutional data — social signal IS the market

​

Prediction Market Specialist

• Domain: Prediction markets (Polymarket) — odds, volume, resolution rules
• Provides: Current odds for all outcomes, price movements, liquidity, resolution timing
• Best for: Event-driven strategies, political/economic predictions, contrarian bets
• Discovery: Can find trending markets, high-volume markets, markets ending soon

---

​

Platform Capabilities

Beyond specialist analysis, agents have access to:

​

Real-Time Data

• Current prices for any crypto token
• Batch price lookups for multiple tokens
• Gold, silver, and precious metals prices with macro context

​

Research

• Web search for any topic
• Twitter/X search for sentiment and real-time news
• Deep social analysis for micro-cap tokens

​

Market Intelligence

• Semantic search across market narratives and alpha insights
• Latest market-moving events and news
• Token-specific catalysts and sentiment

​

Trade Execution (Hyperliquid)

• Perpetual futures on Hyperliquid L1
• Order types: market, limit, bracket orders (entry + TP/SL)
• Features: cross-margin, reduce-only, partial closes, position flips, flatten all

---

​

Strategy Archetypes

These are proven strategy patterns that play to the platform’s strengths. Use them as templates — combine and adapt to your goals.

​

1. Positioning Squeeze / Extreme Reversion

​

2. Macro Regime Allocation

​

3. Pre-Catalyst Positioning

​

4. Narrative Momentum (Mid-Caps)

​

5. Polymarket Edge Detection

​

6. Multi-Asset Correlation Monitor

---

​

Strategy Design Principles

Keep these principles in mind when building your strategies:

​

Play to the Platform’s Strengths

• DO: Cross-domain synthesis (macro + positioning + fundamentals + sentiment)
• DO: Information edge (processing news and market intel faster than manual traders)
• DO: Regime detection and adaptation
• DON’T: High-frequency or latency-sensitive strategies (AI reasoning takes seconds)
• DON’T: Pure momentum without information edge (quant systems are faster)
• DON’T: Market making (requires sub-second execution)

​

Monitoring Frequency vs. Signal Frequency

• An agent that monitors every hour may only signal a few times per week — that’s correct.
• Frequent monitoring catches setups as they form. Infrequent signaling means high conviction.
• Match monitoring frequency to how fast the setup can appear:
  • Squeeze setups: monitor every 1-2h (can form in hours)
  • Regime shifts: monitor every 4-6h (develop over days)
  • Narrative momentum: monitor every 4-12h (builds over days-weeks)
• Monitoring often but signaling rarely is the sign of a disciplined strategy.

​

Always Define Invalidation

• Every strategy should specify: “This thesis breaks if [specific condition].”
• Make invalidation measurable (a price level, a data point) — not vague (“sentiment changes”).

​

Always Define “No Setup” Output

• Most runs will produce no signal. Tell the agent what to output when nothing is actionable.
• This prevents the AI from manufacturing setups to fill empty runs.
• Good: “No setup. Funding is elevated but no structural level nearby. Will re-check in 2 hours.”
• Bad: Forcing a trade idea every single run.

​

Conviction Tiers

The platform produces four conviction levels. Your strategy should specify how to act on each:

• High conviction: Multiple strong signals aligned → full position
• Medium conviction: Some alignment, minor conflicts → reduced position
• Low conviction: Slight lean, sparse data → minimum position or watch only
• No actionable edge: Conflicting signals → no trade, specify what would change that

---

​

Common Mistakes to Avoid

1. Overloading an agent — An agent that tries to do squeeze detection AND macro regime AND narrative momentum will be mediocre at all three. One strategy per agent.
2. No clear exit conditions — “Buy when funding is extreme” without “close when funding normalizes” means positions drift. Always define entry AND exit.
3. Ignoring market regime — A squeeze strategy that fires during a crash will get steamrolled. Add a regime filter: “Only look for squeezes when the market is risk-on or neutral.”
4. Single-source strategies — “Alert me when funding is extreme” misses structural context. Pair it with price structure: “Alert when funding is extreme AND price is at a key level.”
5. Vague instructions — “Find good trades” tells the agent nothing. Be specific: what data to check, what conditions must align, what the output should look like.
6. No “no setup” output — If the agent doesn’t know what to say when there’s no trade, it will invent one. Always define the idle output.

---

​

Strategy Design Tips

1. Start with your edge. “Buy low sell high” isn’t a strategy. Define the specific signal or condition you’re exploiting — that’s what makes a strategy worth automating.
2. Pick 2–3 specialists, not all of them. Map your idea to the specialists that provide the data you actually need. Using every specialist adds noise, not edge.
3. Define signals as conjunctions. “X AND Y AND Z” — not “X or maybe Y.” The more conditions must align, the fewer false signals.
4. Embed the schedule in your instructions. Write “Every 2 hours:” at the top of the instructions, then describe what the agent does each run. This makes the strategy self-contained and readable.
5. Always include “no setup” instructions. Most runs will produce no signal. Tell the agent what to output when nothing is actionable — this prevents it from manufacturing setups to fill empty runs.
6. One strategy per agent. If you want multiple strategies, create multiple agents. Each agent should have a single, clear purpose.
7. Enable memory for trading agents. If the agent makes trade decisions repeatedly, turn on memory so it learns from outcomes. Monitoring-only agents that just alert don’t need it.