Agentic AI coding can embed domain expertise through epistemic grounding documents

As AI-assisted coding moves beyond simple code completion into fully autonomous agent-based software development, a research team at arXiv (cs.SE) is proposing a way to keep domain expertise at the center of the process. The approach uses field-specific epistemic grounding documents — essentially blueprints that encode the hard rules and best practices of a scientific discipline directly into AI agent instructions.

The shift to agentic software development#

The evolution of AI coding tools has accelerated rapidly: from chat-based "vibe coding" to sophisticated agentic systems where human developers sketch a plan and AI agents handle the implementation. This shift democratizes code generation, enabling non-specialists to produce functional software. But speed and accessibility create a risk — without guardrails, generated code may violate disciplinary norms or introduce subtle errors that experts would catch.

Hard constraints and convention parameters#

The proposed grounding document structure encodes two types of rules. Hard Constraints are empirically validated, non-negotiable requirements that ensure scientific correctness — the invariants that must never be violated. Convention Parameters capture community-agreed defaults and best practices that take priority over all other contexts, including user prompts.

The researchers demonstrate the concept using mass spectrometry-based proteomics as a test case, creating a proteomics-specific grounding document that shows how field-scoped rules would translate into AI instructions.

Why this matters#

One critical advantage: AI agents are far more reliable than humans at adhering to explicit guidelines. Unlike developers who might skip a validation step, an agentic system instructed to enforce constraints will do so consistently. This approach allows organizations to preserve domain expert oversight while scaling software generation to non-specialists — embedding best practices at the foundation level rather than relying on review cycles to catch errors.

The model suggests a future where epistemic grounding documents become community resources, kept current by domain experts and governing the behavior of any agentic AI tasked with generating code, tools, or software in that field.