Architecture Case Study · Agentic Systems

AI Control Plane: Markdown-Driven Agent Orchestration in Git

How we orchestrated 21 specialized AI sub-agents with declarative routing tables, durable project memory, and automated verification gates — cutting publishing latency by 70% with 0 build errors.

Vishnu Damwala
9 min read· Updated
Control Plane Summary

The MeshWorld AI Control Plane is a declarative markdown architecture located in the .ai/ directory. It defines deterministic agent routing via .ai/map.md, durable state memory in MEMORY.md, and strictly enforced pre-commit verification scripts for JSON-LD schemas, SEO keywords, and WCAG accessibility.

21

Specialized Sub-Agents

70%

Latency Reduction

0

Production Build Errors

100%

Pre-Commit Test Pass Rate

The Problem: Unconstrained AI Agents in Production Repositories#

LLM-assisted coding and automated content pipelines frequently suffer from prompt drift, hallucinated file paths, breaking changes to schemas, and inconsistent code style.

Instead of treating AI agents as opaque conversational black-boxes, we architected a structured control plane directly inside the Git repository where agents operate as bounded, verified workers with explicit tool boundaries.

Control Plane Directory Specifications (.ai/)#

Mirroring enterprise workflow standards directly in version control

.ai/map.md Routing Matrix

Declarative routing table that maps user intents to a single specialized sub-agent, preventing speculative context pre-loading.

.ai/agent-memory/ Durable Persistence

Session-spanning facts, architectural invariants, and verified domain terms written to durable disk files so agents never re-explore the repository.

.ai/rules/ Path-Gated Behavioral Constraints

Enforced rules (e.g. forbidden buzzwords, JSON-LD schema requirements, exact word counts) activated conditionally based on touched file globs.

Pre-Commit Verification Scripts

Automated deterministic scripts testing SEO links, Acorn JSX safety, and accessibility contrast before changes are staged.

Specialized Sub-Agent Fleet#

Single-responsibility agents with isolated toolkits

.visuals-director

Generates structured infographic prompts, flowchart specs, and alt-text audits.

Validates zero text-spelling errors and proper 16:9 / 1:1 aspect ratios.
.seo-auditor

Validates primary keyword density, H1/H2 hierarchy, and meta descriptions.

Enforces 150-160 character description limits and canonical tags.
.code-reviewer

Checks TypeScript typing, SOLID design rules, and Vitest test coverage.

Fails on missing error boundaries, raw HTML selects, or untyped any.
.aeo-synthesizer

Crafts 40-60 word high-density Quick Answer blocks for AI Overviews.

Validates keyword presence and answer completeness without filler words.

Automated Pre-Commit Verification Gates#

Fail-fast gates eliminating regressions before merge

Link & Anchor Spidering

Live DOM HTTP spider crawler verifying that all internal and external citations resolve (0 broken links).

Acorn JSX & MDX Syntax Audit

Scans markdown problem banks for unescaped curly braces or invalid MDX tags that would break SSR builds.

Structured Schema Verification

Validates Person, WebSite, and TechArticle JSON-LD graphs against schema.org specifications.

Architectural Insights & Takeaways#

1

Markdown is the optimal agent config interface

Human-readable markdown files in .ai/ make agent workflows inspectable, diffable, and editable via standard git pull requests.

2

Deterministic gates beat LLM self-evaluation

Validating output with fast, deterministic TypeScript/Python scripts catches errors that LLMs consistently miss during self-review.

3

Durable memory prevents context thrashing

Maintaining a centralized MEMORY.md file cut token usage by 60% on multi-turn feature implementations.

Frequently Asked Questions (PAA)#

How does an AI Control Plane differ from raw prompts?

Raw prompts are transient and error-prone. An AI Control Plane treats AI agents as formal pipeline components bounded by routing tables, durable state files, and automated verification scripts in Git.

What repositories use this .ai/ control plane standard?

The standard is in production across MeshWorld India (meshworld.in) and ecosystem projects, governing technical writing, content publishing, and code quality.

Can the control plane be adapted to other frameworks?

Yes. The .ai/ specification is framework-agnostic and works with Astro, Next.js, Docusaurus, Vite, and backend microservices.

Explore MeshWorld India in Production

Discover technical articles and architectures built with the AI Control Plane.

Visit MeshWorld Platform
Vishnu Damwala

Vishnu Damwala

Team Lead & Senior Full-Stack Engineer · Technical Architect · Founder at Vishnu Digital

10+ years designing enterprise SaaS, transaction-safe payment architectures, AI agent control planes, and CS educational platforms.

More about the author →
© 2026 Vishnu Damwala. All rights reserved.|Privacy