Skip to content
Profile
E. Burgos
August 27, 2026By Esteban Burgos23:50

Podcast: No more spaghetti code, how SDD-Harness brings order to development with AI

#SpecDrivenDevelopment#SDDHarness#AIAgents#SoftwareEngineering#SoftwareGovernance#CleanCode#WebDevelopment#TechLeadership#AIProgramming
Podcast: No more spaghetti code, how SDD-Harness brings order to development with AI
EpisodePodcast: No more spaghetti code, how SDD-Harness brings order to development with AI
0:0023:50

The Chaos of Uncontrolled Agentic Development

Traditional AI development often relies on direct "trial and error" on the code. The programmer requests changes from the agent, which modifies files destructively, and if something fails, a loop of blind corrections begins.

SDD-Harness stops this behavior by forcing agents to operate under a rigorous engineering framework. Instead of allowing AI to improvise, the kit establishes a flow where software design is the true driver of development, ensuring that every change is predictable, testable, and completely auditable.


The Anatomy of an SDD Cycle: The 7 Specialized Agents

To ensure that software is built with rigor, the kit organizes work into iterative cycles led by a team of 7 specialized agentic agents that act in an ordered sequence of steps:

  1. sdd-orchestrator: The orchestrator. Receives the user's request, evaluates feasibility, initializes the development cycle, and coordinates the rest of the team.
  2. sdd-functional: Analyzes the specification from a business perspective, defining user stories and acceptance criteria in a functional.md file.
  3. sdd-planner: Translates functional requirements into a technical backlog of atomic tasks structured in tasks.json.
  4. sdd-architect: Defines API contracts, data models, and sequence diagrams before coding, ensuring consistency in api.json.
  5. Implementers (Backend / Frontend): Agents dedicated purely to coding. They cannot improvise; their only job is to write code that precisely meets the architect's contracts and the planner's backlog.
  6. sdd-reviewer: The quality inspector. Performs static analysis, runs tests, validates JSON schemas, and calculates the cycle's token telemetry before authorizing closure.
  7. sdd-steward: The kit's concierge. Operates outside the software cycle, handling administrative tasks, general environment health, dependency updates, and repository diagnostics.

The Three Coupling Modes (Zero Friction Integration)

One of the main challenges when adopting a framework is friction in existing infrastructure. SDD-Harness adapts to any project topology through three initialization modes executed with the harness init command:

  • Nx Monorepo: Designed for enterprise-scale architectures that manage multiple applications and shared code libraries under a single modular infrastructure.
  • Standalone: Ideal for independent applications, structuring the harness directly at the root of a single repository in an agile manner.
  • Configure SDD (Legacy/Existing): Via harness configure sdd, the governance kit is injected into existing development without altering a single line of your application's code. The harness is encapsulated in the /sdd folder, and agents inherit and respect the technologies and patterns you were already using.

The 4 Governance Gates: The Quality Shield

The kit's discipline is maintained thanks to four automated quality barriers (Governance Gates) that validate each development stage:

1. SPEC GATE (The Design Contract)

Before an agent can touch a code file, this checkpoint strictly verifies a list of 10 design control points. If any cycle document is missing (such as the technical backlog in tasks.json or the developer's manual approval), development is immediately blocked.

2. CONTEXT GATE (Goodbye to Merge Conflicts)

In environments where multiple developers or agents work in parallel, code conflicts are constant. The Context Gate forces the AI to document its changes through incremental and additive context fragments in Markdown format (sdd/context/.../updates/). These files are directly associated with the active specification cycle, preventing destructive overwriting of global contexts.

3. MEMORY GATE (Continuous Learning)

Each development cycle generates technical learnings. The Memory Gate captures these lessons and records them in session logs (memory/journal/) which are then distilled into a centralized lessons learned file. This way, agents starting new tasks will read this shared memory and will never make the same technical mistake twice.

4. FIX GATE (The Controlled Bypass)

When a critical production issue arises, going through the 7-agent design flow can slow down the solution. The FIX GATE is an agile protocol designed to record quick fixes under specific technical severity prefixes:

  • [HOTFIX]: For critical incidents blocking production (High Severity).
  • [BUGFIX]: For confirmed errors in test environments (Medium Severity).
  • [FIX]: For general corrections.
  • [IMPROVEMENT]: For minor performance or interface improvements (Low Severity).

This bypass has a strict scale limit: if the change affects more than 5 files or alters the central API contract in api.json, the system blocks the bypass and demands starting a standard specification cycle. All fixes are chronologically audited in the sdd/fixes.json index.


Financial Telemetry and the Local Documentation Viewer

For technology leaders (CTOs) and Project Managers, the biggest advantage of SDD-Harness lies in its transparency. By executing:

bash
pnpm sdd:docs

An interactive, bilingual, and 100% offline documentation viewer built purely in vanilla JS is deployed. The viewer silently polls every 4 seconds, updating the status of your tasks, cycles, and fixes in real time without reloading the page or losing your scroll position. Its star feature is the cost telemetry layer. By cross-referencing token consumptions recorded by the sdd-reviewer agent with the per-million token rates configured in sdd/pricing.json (structured with multimodel provider namespaces, e.g., claude/sonnet or gemini/pro), the viewer generates interactive financial charts. These charts realistically contrast the actual agentic cost against the estimated traditional cost based on the human development hours estimated in your backlog, allowing you to audit the return on investment (ROI) of your AI team in real time.


Hermes' Journey: From Idea to Backlog

As a golden touch for the creative flow, the kit includes the Hermes flow (harness idea). By invoking it, the developer can express a business idea in natural language. The system (supported by the sdd-hermes skill and the sdd-steward assistant) will analyze the concept's feasibility, propose an optimal technical stack, structure the initial backlog, and greenlight the SDD cycle under the strict governance of the framework. With SDD-Harness, AI-assisted development stops being an unpredictable black box and becomes what it always should have been: a disciplined, transparent, and high-quality software assembly line.

Written by Esteban Burgos

Comments

Be the first to comment.