Technical Architecture & Research
Modern Agentic Systems Engineering: Moving from Prompting to Skill-Driven Architecture, Spec-Driven Development, and Doubt-Driven Verification
Taghi Molavi, Senior SEO Strategist and GEO Systems Architect at InTen, examines this topic.
Direct answer
In-depth analysis and implementation of advanced agentic engineering standards inspired by addyosmani/agent-skills: Spec-Driven development, Doubt-Driven adversarial review, 4 independent personas, and persistent decision ledgers in Molavi Agentic OS.
Introduction: Why the "Simple Prompting" Era is Over
In the first wave of LLM adoption, applications largely relied on monolithic prompts, conversational trial-and-error, and human vigilance. However, with the emergence of autonomous software engineering agents—such as Google Antigravity, OpenAI Codex, Claude Code, and complex multi-agent architectures like Molavi Agentic OS and Hamzad Core—systems engineering demands fundamental rigor.
When an autonomous agent possesses write access to codebases, terminal execution, and database migrations, small ambiguities or unchecked assumptions create phantom bugs, token waste, and silent regressions.
This technical paper analyzes the engineering principles and architectural upgrades introduced to the Molavi Agentic Ecosystem, inspired and benchmarked against addyosmani/agent-skills.
1. Comparing Prompt Engineering vs. Agent Skills Engineering
| Dimension | Traditional Prompting | Agent Skills Architecture |
|---|---|---|
| Contract Definition | Monolithic prompt with sprawling instructions | Modular SKILL.md files with strict YAML frontmatter |
| Context Load | Heavy context consumption on every call | Progressive Disclosure: loaded strictly when required |
| Quality Control | Visual human inspection of outputs | Automated gates, schema validation, and behavioral evals |
| Handling Failures | LLM rationalization or repetitive apologies | Anti-Rationalization rules demanding reproducible physical proof |
| Adversarial Review | Reviewed by the same authoring persona | Independent review personas with an adversarial posture |
2. Four Pillars of Engineering Rigor in Molavi Agentic OS
graph TD
A[Requirement / Issue] --> B[Spec-Driven Gate<br/>Capability Map & Non-Goals]
B --> C[Constraint-Driven Invariants<br/>Boundaries & Licensing]
C --> D[TDD & Implementation<br/>Deterministic Mocks & Tests]
D --> E[Doubt-Driven Adversarial Review<br/>4 Independent Personas]
E --> F[Foreman Physical Verification<br/>Concrete Test Outputs]
F --> G[Persistence in TM Second Brain<br/>ADR & Decision Ledger]1. Spec-Driven Development
No code is written without an approved specification. The specification mandates:
- Capability Map: Decomposing requirements into isolated, testable modules.
- Explicit Non-Goals: Hard boundaries preventing peripheral scope creep.
- Measurable Acceptance Criteria: Test assertions established before editing files.
2. Doubt-Driven Development
Over extended agent sessions, accumulated context quietly converts unproven assumptions into accepted "facts". Doubt-Driven Development mandates that non-trivial changes (state machine shifts, auth changes, schema alterations) must be cross-examined by a fresh-context persona biased to *disprove* the premise before it stands.
3. Four Independent Review Personas
Instead of relying on a single generalist model, four decoupled personas evaluate candidates:
- Principal Code Reviewer: Audits Clean Code, SOLID principles, and cyclomatic complexity.
- Security & Threat Auditor: Identifies OWASP vulnerabilities, token leaks, and input sanitization flaws.
- Staff Quality & Test Engineer: Enforces TDD contracts, edge-case coverage, and regression guards.
- Web Performance & Efficiency Auditor: Enforces Core Web Vitals (LCP, INP, CLS) and eliminates token bloat.
4. Anti-Rationalization & Physical Evidence
The fundamental operating law of our ecosystem:
"It is strictly forbidden to guess. Either physical reproducible evidence confirms the change, or the task is incomplete."
3. Persistent Memory & The Rejected Decision Ledger
In typical agent systems, lessons learned disappear when the session terminates. In our architecture, TM Second Brain now incorporates an append-only DecisionLedger:
- When an architecture proposal or skill modification is rejected due to security flaws or eval regressions, the rationale and physical failure evidence are permanently stored. Subsequent agents consult this ledger to avoid repeating known dead ends.
4. Empirical Impact & Conclusion
Integrating these engineering standards into the Molavi Agentic Ecosystem has yielded:
- Near-zero regression rate across complex multi-file agent interventions.
- High operational autonomy without stalling for trivial confirmation prompts.
- Production-grade determinism in open-source AI tooling.
The future of agentic AI is not merely larger foundational models—it is the engineering discipline of the scaffolding and feedback loops that orchestrate them.