Human–Agent Consensus Layer
Knowledge as an objective function between humans and agents. The Foundation of Reliable AI-Native Products
When agents are allowed to work autonomously, significant divergence can emerge across the decisions made throughout the implementation process. These divergences may involve design systems, architecture, syntax, product assumptions, or criteria applied without sufficient context.
As more teams adopt agents and prompt them directly for fast outputs, this problem compounds at the speed of the industry itself. Every divergent product or technical decision becomes part of the foundation for the next one. The faster outcomes are demanded and built, the faster these inconsistencies accumulate.
Reducing the Information Delta
Preventing this divergence requires reducing the information delta between:
- your knowledge, prior assumptions, and decisions;
- and the knowledge, assumptions, and decisions held by the agent.
Before any agentic execution moves toward implementation, the human and the agent should reach consensus on every element foundational to the intended outcome: the objective, the relevant context, the underlying assumptions, and the knowledge required to build it correctly.
This consensus should extend even to questions that the human has not yet answered—or has not yet thought to ask. By surfacing and resolving these unknowns before implementation, the agent helps clarify the material intent of the product and enriches the quality of both product thinking and technical execution.
Consensus as an Objective Function
Human–agent consensus can be treated as an objective function.
The agent continuously identifies assumptions, gathers context, tests interpretations, backtests decisions, and refines its priors until its internal model converges with the intended outcome.
Only after this convergence should autonomous implementation begin.
Without it, agents may produce outputs quickly, but those outputs will increasingly encode inconsistent assumptions and compounding technical debt. With it, autonomy becomes trustworthy, repeatable, and scalable.
This is the only scalable way to build reliable products with AI. And you can make these principles consistent across agent memories (CLAUDE.md, MEMORIES.md, etc)
Encode Alignment into the Agent
Do not rely on prompting discipline alone. Encode reusable heuristics and skills that force the agent to inspect existing context, surface hidden assumptions, resolve decisions sequentially, and block execution until shared understanding is confirmed. Matt Pocock’s grill-me, grilling, and grill-with-docs exemplify this pattern: interrogate the intent, ground it in evidence, document the resolved priors, then execute.
Final Reflection
An agent can only scale the product model it has been given. When that model is incomplete, autonomy compounds divergence. When it is shared, explicit, and continuously verified, autonomy compounds intent. The goal is not merely to make agents produce more, but to ensure that every additional unit of agentic work remains faithful to the product you actually mean to build.