Key Outcomes
3-tier
agent hierarchy: company head, department heads, specialized agents
7 phases
structured execution from foundation to deployment
Quality gates
Definition of Done enforced with evidence-based completion
Auditable
change requests, approval chains, and full logging
SDK + IDE
modes for direct-API or IDE-driven execution
Overview
- •
Industry: AI / Developer Tools / Automation
- •
Platform: AI agent orchestration framework
- •
Users: Engineering teams, technical leads, AI agents
- •
Stack: TypeScript · Node.js · Vercel AI SDK · Express · Next.js
Background
Zeus is an intelligent orchestration framework that automates software development with a hierarchical system of AI agents, built for Coverage Creatives to run engineering work as structured, phase-based projects.
Most AI coding tools operate as a single assistant with little structure: no clear ownership, no quality gates, and no audit trail of what changed or why. That makes it hard to trust autonomous output on real projects.
Zeus addresses this by modeling a software company. A company head distributes work to department heads (Architecture, Data, API, UI, and QA & Security), who hand tasks to specialized agents. Every phase has a Definition of Done, changes flow through approval chains, and everything is logged, so execution stays controlled and reviewable.
- •Automate software development through coordinated AI agents, not a single assistant
- •Enforce quality gates and evidence-based completion at every phase
- •Make every change auditable through requests, approvals, and logging
- •Support both direct-API (SDK) and IDE-driven execution modes
Zeus models a software company so AI agents can ship structured, reviewable work instead of unstructured output.
The Challenge
Business Challenges
- Single-assistant AI coding tools lack structure, ownership, and accountability
- Hard to trust autonomous AI output on real, multi-step projects
- No standard process to enforce quality across AI-generated work
- Changes made by AI are difficult to review and audit
Operational Pain Points
- Work isn’t broken down or routed to the right specialization
- No Definition of Done, so completion stays inconsistent
- Errors and ambiguous instructions stall progress without escalation
- No clear record of who changed what, when, and why
Technical Challenges
- Coordinating a three-tier hierarchy of agents across five departments
- Designing seven structured phases with handoffs and quality gates
- Building change management with approval chains and audit trails
- Supporting multiple LLM providers and both SDK and IDE execution modes
The Solution
Zeus is structured as a software company run by AI agents. A company head (Level 0) orchestrates department heads (Level 1) across Architecture, Data, API, UI, and QA & Security, who delegate to specialized agents (Level 2). Projects run through seven phases, from foundation and database through API, UI, integration, testing and security, to deployment, each gated by a Definition of Done and evidence-based completion. A Single Source of Truth drives authority-based permissions, and every change moves through approval chains with full audit trails. Intelligent error handling escalates ambiguous instructions and decomposes failing tasks on a two-strike rule, while logging at company, department, and agent levels keeps each run transparent. Zeus runs in SDK mode (direct LLM calls via the Vercel AI SDK) or IDE mode (the editor acts as the LLM while Zeus manages files and context), and an issue agent turns GitHub issues into structured runs.
Core Architecture
Implementation Process
Architecture & agent design
Defined the three-tier hierarchy, five departments, and specialized agent roles, plus the Single Source of Truth model and authority-based permissions.
Phase & quality framework
Designed the seven execution phases with Definition of Done, evidence-based completion, and a two-strike escalation system.
Change management & logging
Built change requests, approval chains, audit trails, and structured logging at company, department, and agent levels.
Execution modes & integrations
Implemented SDK and IDE modes on the Vercel AI SDK with an Express backend, a Next.js interface, and an issue agent for GitHub-driven runs.
Results & Impact
(Outcomes)Process
- Structured phase-based execution with quality gates
- Auditable changes via approval chains and logging
Coordination
- Hierarchical agents across five departments
- Escalation on ambiguity and repeated failures
Flexibility
- SDK + IDE execution modes on multi-provider LLMs
- Issue-driven runs from GitHub issues to reviewable work
