A developer going by "Anna" on Indie Hackers has published a detailed breakdown of why system prompts alone aren't enough for production AI systems—and she's building NEES Core Engine to solve what she calls 'Agent Drift.' The project, currently in developer preview, adds a governance runtime layer between applications and model providers like OpenAI or Anthropic. Anna is actively seeking 2-3 design partners willing to test the framework on real workflows before broader release.

What Is Agent Drift?

According to Anna's Indie Hackers post, Agent Drift describes how AI agents slowly deviate from intended product behavior once they hit production users. In demos, everything looks polished—but real-world context, business rules, customer pressure, and edge cases cause agents to drift away from their original logic. The failures aren't always dramatic; sometimes the answers sound perfectly reasonable while the underlying decision ignored a policy, skipped a workflow step, or made an untraceable call. The post outlines several concrete failure modes: policy bypass (support bots that sound safe but ignore company terms), pricing hallucination (sales assistants offering unauthorized discounts), context chaos (CRM tools with inconsistent tone based on messy user history), the black box problem (unexplainable decisions), and what Anna calls 'the LLM tax'—repeated expensive model calls for answers that could have been governed, cached, or handled deterministically.

The NEES Architecture

NEES Core Engine positions itself as a governance layer in the request pipeline: App → NEES Governance Runtime → Model Provider → Governed Response. Rather than replacing existing frameworks like LangChain, CrewAI, or model providers, it aims to enforce product-specific rules before responses reach users. The framework includes pre-execution intent checks (understanding user goals before spending tokens), policy enforcement against product-specific rules, memory boundaries controlling what the AI can carry forward, traceable decision logs explaining why responses were allowed or blocked, escalation logic for handing off uncertain cases, cost governance to avoid unnecessary model calls, and fallback behavior during provider failures or latency spikes.

Looking for Design Partners

Anna is explicit that she's not seeking customers—she wants 2-3 founders building AI SaaS products, support agents, CRM assistants, workflow automation tools, internal copilots, education AI, or agentic products with tool use. The goal is to map one real workflow into NEES-style governance structure and measure whether runtime governance reduces Agent Drift in practical environments. Developer preview code is available on GitHub at github.com/NEES-Anna/nees-core-developer-preview, along with a live sample app at naina.nees.cloud.

Key Takeaways

  • System prompts work for demos but lack enforcement power in production AI systems
  • Agent Drift describes gradual deviation from intended behavior across tone, policy, cost, and traceability
  • NEES adds a governance runtime layer between apps and model providers rather than replacing either
  • The creator is seeking design partners to test the framework on real workflows before broader launch

The Bottom Line

This is exactly the kind of infrastructure work that separates hobby projects from production-grade AI systems. If you've been wondering why your 'smart' agent keeps making decisions nobody on your team can explain—or why your token bills keep climbing for repetitive governance tasks—NEES might be worth a look. The developer preview is live and Anna's actively looking for feedback from builders who've hit the system prompt wall.