persona-garden-patch

Task Instruction and Role Specialization as Agent Configuration Layers

Agentic AI systems receive configuration through multiple channels with different scopes and purposes. Without a model for how these channels relate, teams conflate them or treat them as alternatives — using persona files where shared task context belongs, or using shared context files where behavioral specialization is needed.

Two primary channels serve distinct functions that compose into a coherent whole:

Structure

Task Instruction: AGENTS.md

AGENTS.md (emerging across Claude Code, GitHub Copilot, Cursor, and other tools) is a shared, tool-agnostic project context file. It answers: what does an agent need to know to work in this project?

AGENTS.md is read by any tool that supports the convention. It is human-authored and committed to the repository. Its content is stable — it changes when project conventions change, not when an individual agent session starts.

Role Specialization: Agent Persona Files

Agent persona files (agent definitions in platform-specific configuration directories) answer: how should this agent behave in its assigned role?

Persona files are behavioral specifications. They are host-platform-specific: a persona file in one tool’s configuration is not readable by other tools. Persona files change when role definitions evolve — a different cadence from project conventions.

The Orchestrator-Worker Pattern

In multi-agent commission architectures, this two-layer model is instantiated as:

Configuration Layer Content
Project root context file Task instruction Vault conventions, available skills, what not to touch
Agent persona: orchestrator Role specialization Orchestrator thinking, commission management, merge authority
Agent persona: worker Role specialization Execution-level focus, form creation, commit discipline
Agent persona: organizer Role specialization Precinct organization, daily note capture patterns

The project context file tells any agent what the project is and how to work in it. The persona files tell specific role-agents how to approach that work — with different priorities, escalation thresholds, and scope limits.

Boundaries

This model applies to systems where:

The model breaks down when:

The model also does not address the relationship between these two layers and session-specific context — what a single agent carries within an active work session. That is a different layer governed by [[One Context One Concern]]↑ and [[Context Conservation Hierarchy]]↑.

Sources

Relations