Technical Summary: AGENTICAITA

1. Problem Statement

Conventional algorithmic trading systems rely on deterministic heuristics or offline-trained statistical models (e.g., reinforcement learning, static optimization). These systems face a fundamental limitation: they encode knowledge at training time and lack the capacity for explicit, inference-time deliberation when confronted with novel, rapidly shifting market regimes. They cannot adapt to the semantic complexity of new market conditions without retraining, leaving them vulnerable to distributional shifts. Furthermore, existing Large Language Model (LLM) applications in finance (e.g., FinGPT, BloombergGPT) are largely confined to Natural Language Processing (NLP) tasks or advisory roles, failing to integrate directly into the execution control loop with strict safety constraints.

2. Methodology and System Architecture

AGENTICAITA is a proof-of-concept framework designed to replace the traditional "signal-then-execute" paradigm with a fully autonomous, deliberative multi-agent loop. The system operates without offline training or human intervention, relying instead on structured reasoning among specialized LLM agents.

The architecture is deployed as containerized microservices over an isolated private network and consists of four core modules:

A. Adaptive Z-Score Trigger Engine (AZTE)

Acting as a cognitive resource allocator, the AZTE prevents the expensive LLM inference pipeline from activating during low-entropy noise periods.

• Mechanism: It computes an instantaneous return magnitude ($r_t$) and compares it against a rolling baseline (30 observations) to generate an anomaly Z-score ($z_t$).
• Trigger Logic: The pipeline activates if $z_t \ge 2.0$ (statistical anomaly) OR $r_t \ge 0.003$ (absolute return floor).
• Innovation: This dual-condition approach ensures regime invariance, adapting to volatility levels without manual recalibration, while the absolute floor prevents numerical instability during low-volatility periods.

B. Sequential Deliberative Pipeline (SDP)

This is the core agentic contribution, decomposing decision-making into a chain of three specialized agents communicating via typed JSON contracts to prevent role blending and confirmation bias.

1. Analyst Agent: Receives market context (OHLCV, orderbook, funding rates, episodic memory) and the Correlation-Break Diversification (CBD) score. It outputs a structured trading signal (long/short/wait) with confidence and reasoning.
2. Risk Manager Agent: Operates a hybrid gate.
   • Layer A (Deterministic): Enforces hard constraints (e.g., signal validity, confidence $\ge 0.60$, stop-loss limits, max position size of $500) before any LLM call.
   • Layer B (LLM): If Layer A passes, the LLM performs contextual validation and position size calibration.
3. Executor Agent: The only agent with external side effects. It routes orders through a privacy-preserving channel (Tor/VPN) in "LIVE" mode or logs decisions in "DRY_RUN" mode.

C. Inference Gating Protocol (IGP)

A mutex-based scheduler that regulates concurrent access to shared LLM inference resources.

• Function: It serializes agent activations, ensuring that only one pipeline runs at a time. This prevents race conditions, ensures deterministic execution, and enforces a minimum cognitive dwell time, filtering out micro-oscillations.

D. Correlation-Break Diversification (CBD)

A composite score ($\Omega$) integrated into the Analyst's reasoning to operationalize portfolio-level idiosyncratic signal prioritization.

• Formula: $\Omega = \alpha \tilde{z}_t + (1-\alpha)\rho_{cb}$, where $\tilde{z}_t$ is a saturated anomaly score and $\rho_{cb}$ is a decorrelation metric against Bitcoin.
• Purpose: It incentivizes the selection of assets that are moving independently of the broader market, aiming to reduce illusory diversification.

E. Privacy and Execution

The system employs a dual-channel routing architecture: market data is retrieved via direct HTTPS, while order execution is routed through a Tor circuit and VPN to reduce the linkage between agent identity and trading activity.

3. Experimental Results

The framework was validated over a five-day autonomous dry-run session (April 6–11, 2026) on a decentralized perpetual futures exchange using live market data. The system utilized the qwen3.5:9b model via Ollama.

• Operational Correctness: The system executed 157 autonomous pipeline invocations across 76 unique assets with zero human interventions.
• Agentic Friction: The system demonstrated an 11.5% friction rate ($F = 11.5\%$), calculated as the sum of Analyst self-abstentions (8.3%) and Risk Manager hard-gate rejections (3.2%). This confirms that agents exercised independent judgment rather than passively relaying signals.
• Performance Metrics:
  • Net PnL: –$15.07 (–0.058% on $26,079 notional).
  • Win Rate: 51.80% (72 wins, 67 losses).
  • Profit Factor: 0.841 (indicating a structural loss in absolute terms during this specific adverse regime).
  • Benchmark Comparison: Against a passive BTC buy-and-hold strategy over the same period (which lost ~15% or –$3,912), AGENTICAITA achieved a benchmark-relative alpha of +14.94 percentage points.
• Statistical Significance: The authors explicitly state that the sample size (139 trades) is insufficient for statistically robust performance inference. A binomial test yields a p-value of ~0.34, meaning no statistically significant directional edge can be claimed at this stage.

4. Key Contributions

The paper identifies four architectural contributions that distinguish AGENTICAITA from prior systems:

1. Training-Free Orchestration: Unlike RL-based systems, it requires no task-specific training, fine-tuning, or policy optimization, eliminating distributional shift vulnerabilities.
2. Deterministic Safety Constraints: The system enforces verifiable risk limits via hard gates and typed contracts, bounding agent behavior independently of LLM stochasticity.
3. Cognitive Resource Allocation: The AZTE and IGP modules manage computational costs and concurrency, ensuring the system activates only on high-information events and serializes execution for auditability.
4. Episodic Memory & Deliberation: The system implements cross-episode persistence and a structured deliberative pipeline (Analyst $\to$ Risk Manager $\to$ Executor) that allows for inter-agent negotiation and self-abstention.

5. Significance and Claims

The paper positions AGENTICAITA as a preliminary proof-of-concept establishing the feasibility of training-free, deterministic, safety-constrained multi-agent orchestration in financial decision loops.

• Modest Claims: The authors do not claim a statistically robust trading edge or long-term profitability. They explicitly frame the results as "regime-specific operational indicators" demonstrating that the deliberative pipeline functions correctly under live market conditions.
• Primary Value: The significance lies in the architectural validation: proving that a multi-agent system can autonomously perceive, reason, negotiate, and act within strict safety bounds without human intervention.
• Limitations: The paper acknowledges several limitations, including the lack of transaction cost modeling in dry-run (which would reduce PnL further), the long-directional bias of the LLM (90.4% long signals), and the need for a 90-day live deployment with 500+ trades to establish statistical validity.

In summary, AGENTICAITA demonstrates that agentic AI can be engineered to operate safely and autonomously in financial markets, prioritizing operational correctness and auditability over immediate profit generation in this initial validation phase.