Agent Framework

MCP tools, decision framework, and natural language interface for autonomous LLM operation.

The Agent Framework is the interface between the LLM and the fund. It provides the MCP tools for every autonomous action, a decision framework with guardrails, and a natural language interface for complex operations.

MCP Tool Expansion

The existing 44 MCP tools cover trading operations. The autonomous layer adds fund management tools:

Fund Management Tools

Tool	Description
fund_status	NAV, returns, risk metrics, capital available
fund_report	Generate PDF/CSV fund report
allocate_capital	Move capital between strategies
fund_drawdown	Current drawdown vs. limits

Strategy Management Tools

Tool	Description
deploy_strategy	Launch strategy from template + parameters
stop_strategy	Gracefully stop a running strategy
list_strategies	All strategies with health, P&L, state
promote_strategy	Move from staging to live
scale_strategy	Adjust capital allocation
ab_test	Start A/B test between two strategy variants

Research Tools

Tool	Description
scan_opportunities	Scan markets, return ranked by fitness
research_market	Deep analysis of a single market
backtest_hypothesis	Backtest a strategy idea, return results
resolution_status	Check resolution status of active markets
market_universe	Current universe with fitness scores

Risk Tools

Tool	Description
risk_dashboard	Current VaR usage, limits, correlations
stress_test_fund	Run fund-level stress scenarios
risk_budget_status	Per-strategy risk budget utilization
execution_quality	Fill rates, slippage, adverse selection

Explainability Tools

Tool	Description
fund_explain	Full fund state: overview, regime, risk, strategies, decisions, alerts, hypotheses, alpha factors
fund_explain_strategy	Single strategy deep dive: performance, execution, decay, promotion, hypotheses
fund_explain_risk	Full risk analysis: tail risk, Greeks, correlations, stress, VaR, limits, attribution
fund_full_snapshot	Everything in one call

Intelligence Tools

Tool	Description
fund_hypotheses	Active trading hypotheses with lifecycle state, confidence, edge estimates
fund_regime	Current market regime with confidence and recent transitions
fund_alpha_model	Alpha model factor report: per-factor ICs, significance, weights
fund_decisions	Recent autonomous decisions with reasoning, confidence, outcome
fund_decay_report	Alpha decay tracking: edge erosion, half-life, retirement recommendations

Operations Tools

Tool	Description
fund_alerts	Recent alerts with counts by category
fund_ledger	Double-entry journal entries, balance sheet, income statement
fund_promotion_status	Strategy promotion stages and history

Decision Framework

Every autonomous decision goes through guardrails to prevent runaway behavior.

Confidence Thresholds

Different actions require different minimum confidence levels:

Action	Min Confidence	Rationale
Add market to universe	0.3	Low-stakes, just tracking
Deploy strategy (staging)	0.5	Paper mode, no capital at risk
Promote to live (small capital)	0.7	Real money, but limited
Scale up capital	0.8	Increasing exposure
Activate kill switch	0.5	Better safe than sorry
Retire strategy	0.6	Reversible, can redeploy

Escalation Rules

When the LLM should escalate to a human:

Condition	Action
Trade notional > threshold	Route to approval workflow
Confidence < minimum for action	Log decision, wait for human
Unusual market conditions	Alert + pause
Multiple strategy failures	Alert + reduce exposure
Fund drawdown approaching limit	Alert + require human confirmation

Rate Limiting

Prevent runaway decision loops:

# Configure rate limits
decision_limits = DecisionLimits(
    max_deploys_per_hour=3,
    max_capital_changes_per_day=10,
    max_kill_switch_activations_per_day=2,
    cooldown_after_loss=300,  # 5 min cooldown after a strategy is retired
)

Decision Audit Trail

Every autonomous decision is logged with full reasoning:

{
  "decision_id": "d-abc123",
  "timestamp": "2026-03-13T14:30:00Z",
  "action": "deploy_strategy",
  "parameters": {
    "template": "market_maker",
    "market": "will-x-win",
    "capital": 5000
  },
  "reasoning": "Market has 0.87 fitness score, LLM forecast edge of 0.08, backtest Sharpe 1.6 over 45 trades, robustness p=0.02",
  "confidence": 0.75,
  "outcome": "deployed",
  "strategy_id": "s-def456"
}

Natural Language Interface

The LLM translates natural language commands into tool calls:

Natural Language	MCP Tool(s)
“Deploy a market maker on the top 5 political markets”	scan_opportunities -> deploy_strategy x5
”Reduce crypto exposure by 50%“	list_strategies -> scale_strategy for each
”What’s our biggest risk right now?”	risk_dashboard + correlation_matrix
”Why did we lose money yesterday?”	fund_report + execution_quality + decision_log
”Backtest momentum on election markets”	scan_opportunities -> backtest_hypothesis

Dry-Run Mode

For maximum safety, the LLM can operate in dry-run mode where it proposes actions and a human approves them in batches:

# Configure operating mode
agent = AutonomousAgent(
    fund=fund,
    mode="dry_run",  # "dry_run" | "supervised" | "autonomous"
)

# LLM proposes actions
proposals = agent.propose()
# [
#   {"action": "deploy_strategy", "params": {...}, "reasoning": "..."},
#   {"action": "scale_strategy", "params": {...}, "reasoning": "..."},
# ]

# Human reviews and approves
agent.approve(proposals[0].id)
agent.reject(proposals[1].id, reason="too concentrated")

Operating Modes

Mode	Behavior
Dry-run	LLM proposes, human approves each action
Supervised	LLM executes low-risk actions, escalates high-risk
Autonomous	LLM executes all actions within guardrails