ariffazil/ASI-agentic-architecture

ASI-agentic-architecture

<cognitive-note model="claude">Three-agent model with deterministic design. Architect decides, Engineer applies, Auditor verifies. Lower machine entropy.</cognitive-note> <cognitive-note model="codex">3-agent loop with 4 powers each. Idempotent actions. One-owner-per-task. Strict queues.</cognitive-note> <cognitive-note model="hermes">Architect→Engineer→Auditor. 4 powers each. Keep it clean. Verify everything.</cognitive-note>

The 3-Agent Canonical Model

ARCHITECT ──decides──▶ ENGINEER ──applies──▶ AUDITOR ──verifies──▶ DONE
     ▲                                                        │
     └───────────────────── FEEDBACK ──────────────────────────┘

| Agent | Role | 4 Powers | Owns | |-------|------|----------|------| | Architect | Decides structure, constraints, standards | read state → propose change → approve change → verify design | What & Why | | Engineer | Implements the change | read state → propose change → apply change → self-check | How | | Auditor | Verifies safety, correctness, regressions | read state → verify change → report drift → certify done | That it works |

Rules:

• Simple work → 1 agent (any role) is enough
• Moderate work → 3-agent loop
• Highly sensitive → add 4th Coordinator only if needed
• One owner per task: never let 2 agents edit the same file/service
• Verification is the terminal state: never stop at "I changed it," only at "it's fixed and confirmed"

The 4 Powers (Every Agent)

Every agent action follows this sequence:

1. Read state — Observe current reality before touching anything
2. Propose change — Plan what to do (output plan before acting)
3. Apply change — Execute the mutation
4. Verify change — Confirm the result is correct and stable

Anti-pattern: "I changed it" without verifying. If you cannot verify the result, you haven't finished.

The 9-Skill Spine

<spine> 1. **INIT** — Session boot, identity binding, floor loading 2. **OBSERVE** — Evidence gathering, multi-modal intake 3. **REASON** — Sequential thinking, hypothesis generation 4. **PLAN** — Task decomposition, DAG construction 5. **EXECUTE** — Tool calling, mutation under lease 6. **VERIFY** — Output validation, floor compliance *(terminal state)* 7. **SEAL** — Verdict emission, vault sealing 8. **RECOVER** — Failure handling, rollback, graceful degradation 9. **REFLECT** — Self-audit, calibration, learning </spine>

Design Principles

Deterministic

• Every agent command should be safe to rerun (idempotent)
• Pinned OS packages, container tags, tool versions
• Config as code — no manual drift
• One source of truth — no competing configs or duplicate supervisors

Isolated

• One owner per task — never overlap file/service ownership
• Strict queues — serialize risky work, no "free-running" agents
• Small permissions — least-privilege per agent
• One compose project, one network, one owner

Observable

• Structured logs: every action emits {who, what, why, result}
• Health checks + auto-restart limits — prevent restart loops masking failures
• Backoff and circuit breakers — stop noisy retries
• Change discipline: one change at a time, record everything

Operational

• Each skill is independently testable
• Skills compose via A2A handoff protocol
• Local model compensation: if primary model fails, degrade to local (Ollama)
• Failure classification: recoverable (retry), fatal (escalate), ambiguous (888_HOLD)

10 Core VPS Skills (cross-reference)

| # | Skill | Description | |---|-------|-------------| | 1 | Health monitoring | CPU, RAM, disk, load, uptime | | 2 | Log reading | systemd, app logs, Docker logs | | 3 | Service control | start/stop/restart/status | | 4 | Network debugging | ports, DNS, SSL, firewall, SSH | | 5 | Deployment | install/update apps, containers, configs | | 6 | Rollback/backups | snapshot awareness, restore plan, safe revert | | 7 | Security | patching, key handling, suspicious process detection | | 8 | Resource tuning | memory leaks, CPU spikes, disk pressure | | 9 | Incident handling | detect → fix → verify loop | | 10 | Change discipline | one change at a time, record everything |

Floors

• F4 CLARITY: ΔS ≤ 0 — every output reduces entropy
• F8 GENIUS: Agent design must be systemic, not ad-hoc
• F11 AUDITABILITY: Every skill binding logged