agentic-os-kbwen — Agentic OS (KbWen)

Agentic OS by KbWen is a governance-first operating framework for AI coding agents, described as "the governance-first operating system for AI coding agents." It enforces structured workflows through a mandatory phase system (bootstrap → plan → implement → review → test → ship) with hard delivery gates: no evidence means no completion, no gate means no progression. The framework ships 26 slash commands mapped to workflow phases, 14 professional skills (borrowed from the superpowers pattern), plus an .agentcortex/ Single Source of Truth state system with per-task Work Logs. It differentiates by task classification (tiny-fix, quick-win, feature, hotfix, architecture-change) with different required phase sets per classification. A Python-based validation script (validate.sh) checks metadata integrity and command sync at deploy time. The deploy mechanism uses installers/deploy_brain.sh to install the framework non-destructively into any existing project.

Compared to the seeds, Agentic OS is closest to superpowers (Archetype 1 — skills-only behavioral framework) but hybridizes with agent-os (Archetype 4) by adding 26 slash commands on top of the skill layer. The key deltas are: a mandatory gate engine that prevents phase skipping, a SSoT state machine with classification-locked task flows, and a deploy-time installer that works against any existing project rather than requiring a template clone.

Overview — Agentic OS (KbWen)

Origin

Agentic OS v1.1.2 is maintained by KbWen. It positions itself as a governance framework to solve the endemic AI agent problems: skipping steps, hallucinating completion, drifting from scope, losing context, and breaking things silently.

Philosophy

From README:

"No Evidence = No Completion. No Gate = No Progression. No Exceptions."

The framework enforces five principles:

• No Evidence = No Completion — narrative claims are not proof
• Scope Discipline — unauthorized refactoring is strictly prohibited
• Destructive Command Blocking — rm -rf, git reset --hard, force pushes require pre-approved rollback plans
• OWASP Top 10 Auto-Scan — security checks run during /implement and /review
• Confidence Gate — AI must declare confidence level; low confidence triggers escalation

Manifesto-Style Quotes

From AGENTS.md:

"Correctness first. MUST NOT claim completion without verifiable evidence. Small, reversible changes. UNAUTHORIZED REFACTORING STRICTLY PROHIBITED."

"No Bypass Rule: MUST NOT skip Gate/Evidence checks — unknown status = FAIL."

"Read-Once Discipline: Read governance files once at session start; do NOT re-read in later turns."

From README gate diagram:

   Intent          Gate           Workflow         Evidence        Ship
  ┌──────┐      ┌──────┐       ┌──────────┐     ┌──────────┐   ┌──────┐
  │ User │ ───▸ │ Gate │ ───▸  │ Workflow  │ ──▸ │ Evidence │ ─▸│ Ship │
  │ says │      │Engine│       │ + Skills  │     │ Required │   │ SSoT │
  └──────┘      └──────┘       └──────────┘     └──────────┘   └──────┘
                  │ FAIL                           │ FAIL
                  ▼                                ▼
               ⛔ STOP                          ⛔ STOP

Task Classification System

Architecture — Agentic OS (KbWen)

Distribution

• Type: Standalone repository with a deploy installer
• Install (first time):

  git clone https://github.com/KbWen/agentic-os.git
  ./agentic-os/installers/deploy_brain.sh --dry-run /path/to/your-project
  ./agentic-os/installers/deploy_brain.sh /path/to/your-project
  

• Update: bash installers/deploy_brain.sh . (runs from within target project)
• Non-destructive: existing AGENTS.md, CLAUDE.md etc. are NOT overwritten

Directory Tree (after deploy into target project)

<your-project>/
├── AGENTS.md                         # Global governance directives
├── CLAUDE.md                         # Claude Code wrapper
├── .agent/
│   ├── rules/
│   │   ├── engineering_guardrails.md # Classification tiers, gate rules
│   │   ├── security_guardrails.md    # OWASP checks, destructive command blocks
│   │   └── state_machine.md          # Phase transitions
│   └── workflows/
│       ├── bootstrap.md
│       ├── plan.md
│       ├── implement.md
│       ├── review.md
│       ├── test.md
│       └── ship.md
├── .agentcortex/
│   ├── context/
│   │   ├── current_state.md          # SSoT (Single Source of Truth)
│   │   └── work/
│   │       └── <branch-name>.md      # Per-task Work Log
│   ├── bin/
│   │   └── validate.sh               # Deploy-time + CI validation
│   └── docs/
│       └── guides/
│           └── token-governance.md
├── .agents/
│   ├── skills/                       # 14 canonical skills (SKILL.md format)
│   └── workflows/
├── .claude/
│   ├── commands/                     # 26 slash commands
│   └── settings.json
└── installers/
    └── deploy_brain.sh

Required Runtime

• Bash (for installer and validate.sh)
• Python 3.9+ (recommended for full validate.sh — degrades gracefully without it)
• SHA-256 tool (sha256sum, shasum, or openssl)
• Git

Target AI Tools

Claude Code (primary), Cursor, GitHub Copilot, Google Antigravity, Codex — via AGENTS.md + per-agent wrappers.

Components — Agentic OS (KbWen)

Slash Commands (26, in .claude/commands/)

Skills (14, in .agents/skills/)

Validation Script

• .agentcortex/bin/validate.sh — checks metadata integrity, encoding, command sync; runs at deploy time and in CI

State Files

Hooks

None — .claude/settings.json comment confirms: "Sentinel and Phase Summary enforcement is model self-attestation per AGENTS.md — no Python hooks shipped, keeping downstream zero-runtime-dep."

Uniqueness — Agentic OS (KbWen)

Differs From Seeds

Agentic OS is closest to superpowers (Archetype 1 — skills-only behavioral framework) in its 14 professional skills, many of which share names with superpowers skills (test-driven-development, systematic-debugging, using-git-worktrees, verification-before-completion). However, Agentic OS is a hybrid: it adds 26 slash commands on top of the skill layer, creating a command-AND-skill architecture unlike any pure seed archetype. It further hybridizes with agent-os (Archetype 4) in its deploy-into-project installer pattern and SSoT state machine. The distinctive addition is the mandatory gate engine with classification-locked phase chains: where superpowers provides behavioral guidelines, Agentic OS enforces phase gates that the AI literally cannot skip (at the prompt level — "If any gate field is missing, FAIL the gate and list missing fields — do NOT proceed"). The ⚡ ACX end-of-response marker is a session-wide sentinel the validate.sh script can check.

Positioning

This framework occupies a "governance-first" niche between superpowers (pure behavioral) and kiro (IDE-enforced gates). It imposes kiro-like mandatory phase flows without requiring a specialized IDE, relying instead on model self-attestation with explicit failure conditions. The zero-runtime-dependency design choice (no Python hooks) is explicitly documented as a downstream compatibility decision.

Observable Failure Modes

1. Model self-attestation gap: Gate enforcement relies on the model following instructions, not on code-level enforcement. A model that ignores "FAIL the gate" will proceed anyway.
2. Classification lock friction: Reclassification requires rolling back to CLASSIFIED state, which could frustrate users with evolving requirements.
3. Read-Once Discipline violations: The Drift Log requirement for re-reading governance files is easy to forget; token leakage from re-reads is a stated risk.
4. Multi-agent Work Log collision: The "One Branch = One Owner" rule prevents concurrent corruption but requires users to maintain branch discipline.
5. Skill auto-trigger ambiguity: 14 skills that "auto-activate based on task classification and workflow phase" could fire unexpectedly on misclassified tasks.