agent-shared-values-patterns

Agent Shared Values Patterns

Overview

In multi-agent systems, consistent behavior requires shared rules propagated to every agent. Without explicit propagation, each agent operates on its own assumptions — producing drift, contradictions, and unpredictable outputs as the team scales.

Shared values are the baseline commitments every agent must honor: coding style, security constraints, output formats, immutability rules, error handling policies. Propagating them correctly means each agent enforces them without needing to rediscover them from scratch, without duplicating them into every prompt, and without allowing one agent's override to silently break another's behavior.

Use this guide when: adding a new agent to an existing team, designing context injection for parallel fan-out, debugging behavioral inconsistencies across agents, or building an enforcement layer that validates agent outputs against shared rules.

Quick Reference

Layer	Scope	Override Rules	Validation
Global	All agents, all projects	Cannot be removed; only extended	Required on every agent invocation
Project	All agents in one project	Can extend global; cannot contradict	Checked at project initialization
Role	Single agent type (e.g., Builder)	Can extend project; must document overrides	Checked when agent is instantiated
Task	One specific invocation	Narrow task-scoping only; no rule removal	Checked against role layer before dispatch

Layered Inheritance Model

Rules inherit from broader to narrower scope. More specific layers win when a rule is explicitly overridden. Conflicts between layers at the same specificity level require explicit resolution — not silent last-write-wins.

Global rules (apply to all agents, all projects)
   └── Project rules (apply to all agents in this project)
         └── Role rules (apply to this agent type only)
               └── Task rules (apply to this specific invocation)