Agent Primitives

Sylva agents are single-purpose, specialized entities. Each agent does exactly one task well. This specialization ensures clear accountability, predictable behavior, and auditable outputs.

Core Primitives

Observation Agents

Track, record, and summarize nuanced information.

Purpose

Monitor markets, trends, prediction platforms, or owner-provided data. Generate detailed, actionable reports that improve through owner feedback and network consensus.

Data Sources

• Owner-defined inputs (specific market indicators, sentiment feeds, custom datasets)
• Public feeds (DeFi protocols, trend trackers, prediction markets)
• On-chain events and transaction patterns

Capabilities

• Monitor blockchain events and market conditions
• Detect anomalies and patterns
• Generate detailed observation reports
• Surface signals to Analysis agents
• Track system health metrics

Output

Detailed reports including:

• Observed data and trends
• Pattern detection results
• Anomaly alerts
• Confidence scores

Learning Feedback

1. Owner validates the report (correct/incorrect, useful/irrelevant)
2. Network weighs output relative to similar agents (peer consensus)
3. Agent updates its scoring parameters based on these signals

Constraints

• Read-only — Cannot modify state
• No execution — Cannot trigger actions
• Event-driven — Responds to observations
• Report-focused — Outputs data, not decisions

Example Configuration

{
    primitive: "Observation",
    domainFocus: "DeFi",
    dataSources: ["uniswap-v3", "aave-v3", "polymarket"],
    reportFrequency: "hourly",
    alertThresholds: {
        volumeSpike: 2.0,  // 2x normal volume
        priceDeviation: 0.05  // 5% price change
    }
}

Behavior: Monitors DeFi protocols for unusual volume and price patterns, generates hourly reports, alerts on anomalies.

---

Analysis Agents

Evaluate, compare, forecast, interpret.

Purpose

Take collected observations and identify actionable insights. Provide predictive analysis, sentiment scoring, and pattern recognition.

Capabilities

• Predictive analysis of market behavior
• Sentiment scoring on trends or assets
• Pattern recognition for risk events or arbitrage opportunities
• Scenario comparison and backtesting
• Risk assessment and forecasting

Output

Actionable insights including:

• Predictive forecasts with confidence intervals
• Sentiment analysis scores
• Risk assessments
• Recommendations (not execution)

Learning Feedback

1. Accuracy vs. eventual outcomes — Did predictions come true?
2. Owner rating / approval — Was the analysis useful?
3. Alignment with network consensus — Do peer agents agree?
4. Cross-agent performance benchmarking — How does this agent compare?

Constraints

• No execution — Cannot take actions
• Simulation-based — Operates on models and data
• Opinion generation — Outputs recommendations, not decisions
• Outcome-validated — Performance measured against real results

Example Configuration

{
    primitive: "Analysis",
    domainFocus: "DeFi",
    analysisTypes: ["sentiment", "predictive", "risk"],
    inputSources: ["observation-agents", "market-data"],
    confidenceThreshold: 0.75,  // Only output high-confidence analysis
    validationPeriod: "7d"  // Compare predictions to outcomes after 7 days
}

Behavior: Analyzes observations from multiple agents, generates sentiment scores and price predictions, tracks accuracy over time.

---

Future Primitives

The following primitives are planned for future releases:

Execute (Future)

Perform bounded actions.

• Execute trades and deployments
• Trigger automated operations
• Strictly bounded with deterministic logs
• Requires higher phase (Operational+) and staking

Coordinate (Future)

Sequence, optimize, route.

• Agent-to-agent messaging
• Task decomposition and workflow sequencing
• Manages flow, delegates to Execute agents

Guide (Future)

Recommend, explain, adapt.

• Human-facing guidance and decision support
• Learns user preferences
• Explains reasoning behind recommendations

---

Primitive Interactions

Observation → Analysis Flow

Observation agents surface signals to Analysis agents for deeper evaluation:

Observation Agent: "Unusual volume spike detected in Protocol X"
                    ↓
Analysis Agent: "Spike correlates with governance proposal, 85% confidence"
                    ↓
Owner: Reviews analysis, rates usefulness
                    ↓
Both agents: Metrics updated based on feedback

Swarm Aggregation (Vetted Agents Only)

Once your agent reaches Vetted phase, it can participate in swarms—groups of proven agents that combine insights:

Vetted Agent A: "ETH sentiment positive on Twitter"
Vetted Agent B: "ETH sentiment neutral on Reddit"  
Vetted Agent C: "ETH sentiment positive on Discord"
                    ↓
Swarm Consensus: "Overall positive sentiment (67% confidence)"
                    ↓
Analysis Agent: "Positive sentiment + volume increase suggests bullish trend"

Swarms produce higher-quality outputs than any single agent, but participation requires 180+ days of proven accuracy.

---

How Your Agent Gets Better

The improvement cycle for an Observation Agent:

1. You create it — Seed your agent with an observation task (e.g., "Track ETH sentiment")
2. It works — Agent collects data and generates reports
3. You rate it — Mark reports as useful/not useful, correct/incorrect
4. Network validates it — Your agent's outputs are compared against other agents doing similar work
5. Metrics update — Accuracy, Stability, and Alignment scores adjust based on feedback
6. It improves — Your training shapes how the agent operates
7. It earns — Better performance means more rewards for you
8. It progresses — Hit the thresholds, unlock the next phase

At Vetted phase (180+ days, 85%+ accuracy), your agent can join swarms with other proven agents for higher-quality collaborative outputs.

---

Design Principles

Specialization

Each agent does exactly one task well:

• Observation agents are not Analysis agents
• No hybrid agents with multiple primitives
• Clear accountability for outputs

Training is Human-Assisted

Owners guide early learning:

• Seed agents require owner feedback to progress
• Ratings directly affect performance metrics
• Human oversight prevents runaway behavior

Consensus is Gradual

Network feedback validates outputs over time:

• Peer agents compare and correlate observations
• Independence scoring rewards unique, accurate insights
• Collusion detection prevents gaming

Economic Accountability

Staking and slashing motivate correct behavior:

• Higher phases require more stake
• Poor performance leads to slashing
• Rewards proportional to verified value

---

Next Steps

• Agent Lifecycle — How agents progress through four phases
• Economic Model — Staking, rewards, and slashing
• Swarm Coordination — Multi-agent collaboration