AgenticWorkflow: Universal Agentic Skill & Execution Toolchain

A pluripotent stem-cell framework and multi-agent execution engine that turns complex tasks into deterministic, self-verifying autonomous workflows.

Execution Invariant

$$\text{Intent} \xrightarrow{\text{Research}} \text{Plan (SOT state.yaml)} \xrightarrow{\text{Implementation}} \text{4-Layer Quality Gates} \xrightarrow{\text{pACS Delta}} \text{Verified Deliverable}$$

⚡ Dynamic Mode Router

Detect the desired mode from context or explicit flags:

3 Absolute Criteria (Constitutional Canon)

Every workflow, tool, and subagent created or governed by this skill strictly inherits the 3 Absolute Criteria:

1. Absolute Criterion 1: Quality of the Final Deliverable

   Speed, token cost, workload, and length limits are completely ignored. The sole criterion for every decision is the quality of the final deliverable.

2. Absolute Criterion 2: Single-File SOT + Hierarchical Memory

   All shared state is concentrated in a single file (state.yaml). SOT write permission belongs exclusively to the Orchestrator / Team Lead. Parallel agents never mutate shared files simultaneously.

3. Absolute Criterion 3: Code Change Protocol (CCP)

   Before writing, modifying, adding, or deleting code, internally perform: Step 1 (Understand Intent) → Step 2 (Ripple Effect Analysis) → Step 3 (Change Plan). Governed by Coding Anchor Points (CAP-1~4).

4-Layer Quality Assurance Stack

flowchart TD
    StepStart["Step Execution"] --> L0["L0: Anti-Skip Physical Guard (File exists & >= 100 bytes)"]
    L0 -->|PASS| L1["L1: Verification Gate (100% functional goal achievement)"]
    L0 -->|FAIL| Diag["Abductive Diagnosis"]
    L1 -->|PASS| L15["L1.5: pACS Self-Rating (F/C/L Pre-mortem scoring)"]
    L1 -->|FAIL| Diag
    L15 -->|RED: <50| Diag
    L15 -->|GREEN / YELLOW| L2["L2: Adversarial Review (@reviewer + @fact-checker)"]
    L2 -->|PASS| SOT["Advance SOT (current_step + 1)"]
    L2 -->|FAIL| Diag
    Diag --> Retry["Retry with Alternative Hypothesis (Max 3)"]

1. L0 Anti-Skip Guard: Deterministically verifies deliverable exists on disk and is non-empty (MIN_OUTPUT_SIZE >= 100 bytes).
2. L1 Verification Gate: Semantic verification that all acceptance criteria are 100% achieved.
3. L1.5 pACS (Predicted Agent Confidence Score): Pre-mortem evaluation across Faithfulness, Completeness, Logic. $pACS = \min(F, C, L)$.
   • GREEN (>= 70): Auto-advance.
   • YELLOW (50 - 69): Flag weak dimension in Decision Log and proceed.
   • RED (< 50): Halt and trigger rework.
4. L2 Adversarial Review: Independent Generator-Critic evaluation by @reviewer and @fact-checker.

Autonomous Self-Fueling Engine & Circuit Breaker

The system operates as an end-to-end autonomous engine with built-in energy management:

1. Self-Fueling & Energy Loop: Dynamic token and context headroom monitoring with automatic RLM state compaction and refueling checkpoints before context exhaustion.
2. Fable Circuit Breaker: State transitions CLOSED → OPEN → HALF_OPEN. Automatically halts speculative thrashing when consecutive failure streak $\ge 2$, isolates root causes via Abductive Diagnosis (diagnose_context.py), and executes structured recovery.
3. Clean Code Guard: Automated AST guard pass checking the 24 Clean Code imperatives (small functions, intent-revealing names, maximum 4 parameters, no swallowed exceptions, no fake success mocks).
4. AI Engineering Evaluation: Integrated four-fifths disparate impact testing ($\ge 0.80$), PSI distribution drift monitoring, and adversarial prompt-injection sanitization.
5. OmniSkill Dynamic Intent Routing: Decomposes complex natural language intent into an optimal multi-step DAG across Research, Planning, Implementation, and Verification.

CLI & Toolchain Integration

# Launch autonomous end-to-end autopilot workflow with self-fueling & energy management
agentic-workflow autopilot --title "Production Pipeline" --goal "Autonomous Delivery"

# Run Clean Code Guard audit pass (SOLID, 24 Imperatives, AI failure modes)
agentic-workflow guard [directory]

# Execute AI Engineer fairness, drift, and prompt-injection evaluation gates
agentic-workflow eval

# Query multi-agent observable trace logs and spans
agentic-workflow traces

# Run OmniSkill dynamic intent routing and multi-host validation
agentic-workflow omni-skill route "build an autonomous pipeline"
agentic-workflow omni-skill validate

# Check live workflow dashboard and observability metrics
agentic-workflow status

# Run full automated test suite (safety, guard, MAS, evaluator)
agentic-workflow test

Universal Multi-Agent & Multi-Host Distribution

Install and use AgenticWorkflow instantly across any platform or AI harness:

# Skills.sh / Vercel AI (Universal Agent Discovery)
npx skills add imMamdouhaboammar/agentic-workflow

# Zero-Install CLI (Bun & Node/npx)
bunx agentic-workflow --help
npx agentic-workflow --help

# Python Environment (pip / PyPI)
pip install agentic-workflow

# Universal Multi-Agent Installer (Claude, Gemini, Cursor, Codex, OpenCode)
curl -fsSL https://raw.githubusercontent.com/imMamdouhaboammar/agentic-workflow/main/install.sh | bash