Technical Documentation

Agent Orchestration &
LLM Routing Architecture

A deep dive into how Nova Core coordinates 14 specialised AI agents, routes tasks to the right LLM, and maintains a governed, auditable SDLC pipeline.

01Architecture Overview

Nova Core is built on a multi-layer agentic architecture. At its core is the Nova Orchestrator — a stateful coordination engine that manages agent lifecycles, task graphs, message passing, and HITL approval gates. Every agent is a microservice-like entity: stateless, LLM-backed, and scoped to a single domain.

CLIENT INTAKE INTERFACE NOVA ORCHESTRATOR Task Graph Engine State Manager Message Bus HITL GATEWAY Approval Queue LLM ROUTER Model Selector AGENT POOL BRD Agent Spec Agent Architect Agent Scrum Agent UX Agent Frontend Agent Backend Agent Cloud Agent DB Agent Security Agent QA Agent Release Agent Support Agent CONTEXT STORE Vector DB + Short-term Memory INTEGRATIONS Jira · GitHub · Slack LLM MODELS Claude · GPT-4o · Gemini

02The Nova Core Orchestrator

The Orchestrator is the central nervous system of Nova Core. It maintains a directed acyclic task graph (DAG) for each project, where each node represents an agent task and each edge a dependency. The orchestrator schedules tasks based on graph topology, monitors completion, handles retries, and triggers HITL gates at defined phase boundaries.

The Orchestrator is stateful and survives restarts. All task state, agent outputs, and approval decisions are persisted in the Nova Core state store and are fully auditable.
Core Responsibilities

Task Graph Management — Build, validate, and traverse project-specific DAGs derived from the LLD document
Agent Lifecycle — Spawn, monitor, pause, and terminate agents based on graph state
Message Bus — Route inter-agent messages and shared context payloads via an async pub/sub system
HITL Gate Enforcement — Block downstream tasks until human approval is received at defined checkpoints
Retry & Fallback — Detect agent failures and re-route to secondary LLM or escalate to human review

// Orchestrator task dispatch — simplified async function dispatchTask(task: AgentTask) { // Resolve dependencies from task graph const deps = await graph.getDependencies(task.id); if (!deps.every(d => d.status === 'completed')) return; // Check for HITL gate if (task.requiresApproval) { await hitlGateway.requestApproval(task); return; // Paused until human approves } // Route to correct agent with LLM assignment const agent = agentRegistry.get(task.agentRole); const model = llmRouter.select(task.agentRole, task.complexity); await agent.execute({ ...task, model }); }

03Agent Registry & Role Definitions

Each agent is registered in the Agent Registry with a role schema that defines its input/output contracts, allowed LLM models, token budgets, HITL requirements, and integration hooks.

AgentPrimary ModelToken BudgetOutput TypeHITL
brd-agentclaude-opus-432kStructured JSON + MarkdownRequired
spec-agentclaude-opus-448kFSD DocumentRequired
arch-agentclaude-opus-464kHLD + LLD + DiagramsRequired
scrum-agentgpt-4o16kJira Payload (REST)Required
ux-agentclaude-sonnet-424kWireframe Specs + TokensOptional
frontend-agentclaude-sonnet-480kGit Commits (PR)Optional
backend-agentclaude-sonnet-496kGit Commits (PR)Optional
cloud-agentclaude-sonnet-448kTerraform / CDK FilesOptional
security-agentgpt-4o32kAudit Report JSONRequired
qa-agentclaude-sonnet-464kTest Suites + ReportRequired
release-agentgpt-4o-mini16kCI/CD Config + NotesRequired

04LLM Routing Engine

The LLM Routing Engine dynamically selects the optimal language model for each agent task based on four signals: task complexity, output type, latency requirements, and cost budget. This ensures high-stakes tasks (architecture, security) use the most capable model while routine tasks are handled cost-efficiently.

Routing Decision Matrix
Task CategoryPrimary ModelFallbackReason
Strategic / DesignClaude Opus 4GPT-4oLong-context reasoning, nuanced output
Code GenerationClaude Sonnet 4GPT-4oCode quality + speed balance
Structured DataGPT-4oGemini 1.5 ProJSON output reliability
Fast / LightweightGPT-4o miniGemini FlashCost efficiency, low latency
Multi-modal (UX)Claude Sonnet 4Gemini 1.5 ProVision + text for wireframe interpretation
// LLM Router — model selection logic function selectModel(role: AgentRole, task: Task): LLMConfig { const complexity = analyseComplexity(task.context); const budget = task.tokenBudget; if (STRATEGIC_ROLES.includes(role) && complexity > 0.7) { return { provider: 'anthropic', model: 'claude-opus-4-5' }; } if (CODE_ROLES.includes(role)) { return { provider: 'anthropic', model: 'claude-sonnet-4-5' }; } if (budget < 8000) { return { provider: 'openai', model: 'gpt-4o-mini' }; } return { provider: 'openai', model: 'gpt-4o' }; }

05Human-in-the-Loop (HITL) Gateway

Every major phase boundary in the pipeline has a mandatory HITL gate. The gateway sends a structured notification to the project stakeholder via the Nova Core Dashboard (and optionally Slack/email). Stakeholders can Approve, Request Changes, or Reject with inline comments. Agents remain paused until a decision is received or the SLA timeout triggers an escalation.

HITL gates cannot be disabled at the BRD, Architecture, and Release phases. This is a hard constraint built into the orchestrator to ensure human oversight on all critical deliverables.

06Context & Memory Management

Nova Core maintains a two-tier memory system per project. The Short-Term Context Window carries the last N agent interactions and is passed as system context to active agents. The Long-Term Vector Store persists all documents, decisions, and code artifacts, allowing any agent to semantic-search project history at any point in the lifecycle.

07Parallel Execution Engine

Once the Architecture phase is complete and approved, the Build phase spawns up to 5 concurrent agents (Frontend, Backend, Cloud, Database, AI Stack). Each agent works on isolated branches and communicates shared contracts via the API specification locked in the LLD. The Orchestrator monitors for conflicts and mediates via a shared contract registry.

08Security & Audit Trail

Every agent action — including prompts sent, model responses received, decisions made, and tools called — is written to an append-only audit log. This log is immutable, tamper-evident, and exportable. The Security Agent performs a continuous OWASP scan on all generated code and blocks release if critical vulnerabilities are found.

Nova Core is designed for SOC 2 Type II compliance. All data is encrypted at rest (AES-256) and in transit (TLS 1.3). Customer LLM API keys are stored in an isolated secrets vault and never logged.

Nova Core Technical Docs v1.0 — April 2026 — Confidential

Now in Beta — 500+ teams building smarter

A Constel Global Product

Your Entire Engineering Team, Powered by AI.

Nova Core deploys a full virtual engineering organisation — from requirements to release — using specialised AI agents that work in parallel, 24 hours a day.

Start Building Free See How It Works
14+
Agent Roles
24/7
Always Building
60%
Faster Delivery
100%
Human Approved
scroll
The Problem

Building software shouldn't need
an army of engineers

Traditional software development is broken — slow, expensive, and talent-dependent.

Months of Hiring
Building a cross-functional team takes 3–6 months and $1.5M+ per year in salaries alone.
Coordination Chaos
BA, UX, dev, QA, cloud — keeping every team aligned burns weeks of leadership time.
Quality Gaps
Context is lost between handoffs. Docs go stale. Deadlines slip. Quality suffers.
Runaway Costs
45% of software projects run over budget. Engineering overhead eats product investment.
The Solution

Meet your AI engineering
organisation

A constellation of 14 specialised AI agents — each an expert in their domain — working in parallel, 24 hours a day.

How It Works

From idea to production
in a single pipeline

Every step AI-powered. Every gate human-approved. Every deliverable production-grade.

Agent Roster

Every role. Every phase.
Every time.

14 specialised AI agents covering your entire SDLC — from first idea to production support.

AgentReal-World RoleKey DeliverablePhase
Pricing

Simple pricing.
Launching soon.

Crafting pricing that works for every stage of growth — from solo founders to enterprises.

Proudly built by — Innovation Division
Get Started Today

Ship your product.
Without the overhead.

Join 500+ founders and product teams using Nova Core to build faster, smarter, and leaner.