This document formalizes the division of labor between the Human Developer (the Epistemic Agent) and the Coding Agent (the Stochastic Pattern-Completion System) based on the structural differences in their epistemic pipelines.
The Human Developer is responsible for judgment—the world-directed, value-infused commitment to a solution. The Coding Agent is responsible for linguistic automation—the efficient synthesis of statistically plausible code patterns.
| Pipeline Stage | Coding Agent Responsibility (High-Volume Synthesis) | Human Developer Responsibility (High-Value Judgment) |
|---|---|---|
| 1. Grounding | Textual Processing: Receiving requirements and existing codebases as symbol sequences. | World Context: Providing the sensorimotor, social, and business reality that text alone cannot capture. |
| 2. Parsing | Tokenization: Mechanically segmenting input into efficient numerical representations. | Situational Parsing: Interpreting the "why" behind the code and identifying objects, agents, and goals. |
| 3. Experience | Statistical Association: Leveraging clusters of how words/code co-occur across massive datasets. | Episodic Memory: Applying past lived experience, intuitive physics, and specific architectural history. |
| 4. Motivation | Inference Optimization: Minimizing predictive error to generate the most likely continuation. | Goal Setting: Defining what "success" or "safety" means for the specific project and its users. |
| 5. Causality | Context Integration: Maintaining thematic coherence and tracking referents within the code. | Causal Reasoning: Constructing the logic of how changes will affect the whole system under stress. |
| 6. Metacognition | Continuous Production: Always providing a response to satisfy the statistical trajectory. | Error Monitoring: Identifying hallucinations and withholding judgment when the agent is "confidently wrong". |
| 7. Value | Probabilistic Prediction: Delivering the most likely code sequence based on training data. | Value-Sensitive Judgment: Taking accountability for the moral, legal, and operational consequences of the code. |
- Anti-Epistemia Check: The developer must never treat "fluent" or "correct-looking" code as "justified" without manual verification.
- The Labor of Judgment: The developer must consciously perform the "cognitive labor" of understanding the logic, rather than just consuming the generated product.
- Process-Sensitive Probes: Instead of just checking if the code runs, the developer must test for causal stability—asking if the code tracks interventions or merely surface associations.
- Preservation of Accountability: Responsibility for errors remains exclusively with the human, as the agent lacks the internal machinery for belief, commitment, or stakes.