AGENTIC_HARNESS

Agentic Harness

Three-layer architecture for AI-assisted development at agentic-workstation.

---

Overview

The agentic harness is a framework for structuring AI agent interactions. It defines how agents discover context, make decisions, and deliver work — across any AI tool (Claude Code, OpenCode, Cursor, Windsurf).

---

Three layers

Layer	Name	Purpose
L1	Infrastructure	Workstation bootstrap, skills, agents (this repo)
L2	Session	Running workspace, knowledge base, personas, packs
L3	Delivery	Workflow execution, human gates, PR creation

---

Session lifecycle

1. Init — load workspace context, inject knowledge
2. Discovery — scan repo docs in fixed order (README → docs → AGENTS.md → ...)
3. Plan — break down work, identify risks
4. Execute — implement with quality checks
5. Review — self-review + human gate
6. Deliver — create PR, update tickets

---

Personas

Personas constrain what an agent is allowed to do:

Persona	Constraint
implementer	Write code, bias toward action
reviewer	Analyze and critique, no changes
researcher	Explore and summarize, no implementation
architect	System design, tradeoffs, ADRs

---

Packs

Packs bundle context for a specific client or project:

• Repository paths and remote URLs
• Project/ticket IDs and conventions
• Tool-specific overrides
• Default automation level

---

Key principles

• Document-first: Always read docs before code
• AGENTS.md as contract: Primary source of truth for agent behavior
• Human gates: Required between phases for non-trivial work
• English artifacts: All commits, PRs, and tickets in English

---

Canonical doc: docs/AGENTIC_HARNESS.md