daVinci-Agency Paradigm

• daVinci-Agency Paradigm is a framework for designing agentic AI systems that preserve human agency while enabling long-horizon, goal-directed workflows.
• It employs hierarchical multi-agent architectures with orchestration, task-specific agents, and PR chain mining to achieve modular and robust performance.
• The paradigm emphasizes quantitative performance metrics and agency-safe design principles to ensure reliable, human-centric AI collaboration in complex environments.

The daVinci-Agency paradigm is a set of methodologies and frameworks for designing agentic artificial intelligence systems that excel in long-horizon, goal-directed workflows while rigorously preserving human agency and ensuring robust, modular decision-making. Originating in both surgical human–AI teaming and software development, daVinci-Agency reframes agent coordination, supervision signal extraction, and safety guarantees to address complex, multi-stage environments. The paradigm tightly integrates hierarchical multi-agent orchestration, agency-preserving objectives, long-term consistency constraints, and authentic supervisory data derived from real-world project evolution (Park et al., 10 Nov 2025, Mitelut et al., 2023, Jiang et al., 2 Feb 2026).

1. Foundational Principles and Formal Definitions

At the core of daVinci-Agency is an explicit commitment to agency preservation in AI–human interaction. Agency is mathematically defined as $agency(s) = \lvert A(s) \rvert$, where $A(s)$ is the set of actions available to the human at state $s$. The forward-looking agency metric evaluates accumulated future option breadth:

$$E^{\pi}(s) = \mathbb{E}\left[\sum_{t=0}^\infty \gamma^t\,agency(s_t)\right], \quad 0 \leq \gamma < 1$$

For joint human–AI policies $\pi_\text{sys}$ and a human-alone baseline $\pi_H$, interactions are agency-preserving if, for all reachable states $s$,

$E^{\pi_\text{sys}}(s) \geq E^{\pi_H}(s)$

This formalism underpins decision-making rules for both software agents and safety-critical systems, ensuring that the cumulative capacity for user-directed choice is never diminished by AI intervention (Mitelut et al., 2023). Deviations at each timestep are measured via the “agency-error” $\delta_t^A$, analogous to TD error in reinforcement learning.

2. Hierarchical Multi-Agent Architectures

The paradigm operationalizes agent orchestration through a layered architecture, exemplified by the Surgical Agent Orchestration Platform (SAOP):

• Orchestration Agent: Implements high-level workflow functions with modular subcomponents:
  • Profile Module: Encodes top-level objectives (e.g., valid wake-word detection, workflow boundaries, agent plug-and-play).
  • Memory Module: Tracks command history and current state, spanning global (across video clips) and local (within a clip) memory.
  • Decision Module: Leverages LLMs (e.g., Gemma3 with chain-of-thought prompting) to select workflow actions via a probabilistic JSON API, applying argmax over per-function probabilities.
  • Feedback Module: Monitors invalid cycles, restricts re-entries to a strict upper bound (≤3), and enforces clarity in speech-command dialogue.
• Task-Specific Agents:
  • Information Retrieval (IR): Extracts patient clinical data for overlay, making field selection decisions via LLM-inferred distributions $\pi_\theta(a|x,i)$ and $A(s)$0.
  • Image Viewer (IV): Navigates and zooms through CT scan volumes; state transitions are governed by policy factorization and deterministic action–parameter selection.
  • Anatomy Rendering (AR): Manipulates 3D anatomical reconstructions with continuous state parameters for zoom, rotation, and structure visibility, ensuring temporally coherent action via smooth interpolation.

This architectural decoupling ensures both robustness and extensibility, supporting new data modalities and workflow configurations without cross-component re-engineering (Park et al., 10 Nov 2025).

3. Long-Horizon Agentic Supervision via Real-World Software Evolution

daVinci-Agency reconceptualizes agentic output supervision by mining Pull Request (PR) chains from mature software repositories. Each PR chain, $A(s)$1 with $A(s)$2, encodes:

• Progressive Task Decomposition: Serial commits naturally decompose project-scale objectives into granular, verifiable actions.
• Long-Term Consistency Enforcement: Cross-PR consistency with the global project goal $A(s)$3 is enforced through embedding-based similarity constraints and temporal drift regularization.
• Verifiable Refinement: Authentic bug-fix pairs $A(s)$4 allow targeted supervision on the refinement and debugging subskills via performance delta metrics $A(s)$5.

Agentic policies $A(s)$6 are trained to jointly minimize three losses—$A(s)$7 for decomposition, $A(s)$8 for consistency, and $A(s)$9 for refinement—weighted to promote robust long-horizon task execution (Jiang et al., 2 Feb 2026).

4. Evaluation and Quantitative Performance

Multi-Level Orchestration Evaluation Metric (MOEM)

The MOEM framework provides rigorous assessment of the orchestration pipeline:

• Stage-Level Accuracy: For each stage $s$0, $s$1
• Workflow-Level Success: Strict, single-pass, and multi-pass rates ($s$2, etc.) evaluated over complete end-to-end flows
• Category-Level Success: $s$3 computed for command structural, type, and expression categories

Representative results in a voice-controlled surgical environment:

Stage	Accuracy or Success Rate
Real-Time Audio (wake)	88.6%
STT transcription	~81% (raw), CC ≈97%
Command Reasoning	≈98%
Action Determination	≈95%
Orchestration Flow	100%
SR_multi-pass	95.8%

Category-level robustness metrics confirm high performance even under paraphrased and composite command input, with failure cases concentrated in composite+paraphrase intersections (Park et al., 10 Nov 2025).

Software Agentic Workflows

With only 239 curated PR chains (average 85,000 tokens, 116 tool calls), models trained under the daVinci-Agency paradigm achieve a 47% relative gain on the Toolathlon benchmark and robust improvements across SWE-bench, DS-1000, $s$4-Bench, and SciCode-MP. Token and tool call efficiency scales favorably relative to synthetic or flat-trajectory baselines (Jiang et al., 2 Feb 2026).

5. Agency-Safe AI Design Principles

To guarantee agency preservation and robust performance, four design pillars are articulated (Mitelut et al., 2023):

1. Explicit Agency-Augmented Objective: Reward functions include an explicit forward-looking agency term; $s$5.
2. Baseline Agency Tracking & Auditing: Continuous logging and comparison of $s$6 vs. $s$7, enabling post-hoc verification and regulatory compliance.
3. Transparent, Choice-Preserving Interfaces: UI presents alternative actions and explicit “agency scores” for each trajectory.
4. Conservative Agency Policy Under Uncertainty: Under model uncertainty, the agent defaults to policies that preserve maximal worst-case agency, solving a robust MDP for conservative deployment.

6. Data Mining and Practical Implementation Pipeline

The structured collection of long-horizon supervision proceeds via:

• Repo Selection: ~61,800 PRs mined from repositories such as numpy, apache/pulsar, scipy.
• Dependency-Aware PR-Chain Construction: Construction of directed acyclic PR dependency graphs to extract multi-PR chains.
• Query Synthesis: Sub-queries for each PR, omitting implementation but conveying change intent for semantic alignment.
• Rollout and Rejection Sampling: GLM-4.6-based rollouts, state carryover, and strict acceptance thresholds ($s$8) for dataset quality.

This pipeline underpins the efficient acquisition of authentic, causally-structured agentic trajectories for LLM fine-tuning (Jiang et al., 2 Feb 2026).

7. Broader Implications and Future Directions

The daVinci-Agency paradigm supports multi-domain generalization. In surgery, modular orchestration yields robustness to diverse command expressions and plug-and-play modality extension (e.g., intraoperative ultrasound), while data-efficient long-horizon learning in software unlocks persistent, project-scale goal pursuit.

Outlined future directions include:

• Integration of agency-preserving frameworks into RL from internal states, tracking variables such as user confidence and cognitive load.
• Formalization of algorithmic human rights within AI policy spaces.
• Mechanistic interpretability of agency within neural architectures, enabling third-party auditing.
• Extension to multilingual and mixed-modality command environments, necessitating advances in STT adaptation and transliteration.

A plausible implication is that wide adoption of daVinci-Agency’s principles can mitigate agency erosion in both supervised and reinforcement learning contexts, promoting human–AI collaboration where user autonomy and long-term options are preserved or enhanced.