Skip to content
/

AI Agents in Enterprise Product Development Spec

enterprise-ai-agents-spec · last commit 2025-07-19

Primitive shape
No installable primitives
00

Summary

enterprise-ai-agents-spec — Summary

enterprise-ai-agents-spec (51 stars) is an open-source methodology document authored by Zohar Babin providing a comprehensive blueprint for coordinating AI agent swarms in enterprise product development. The primary artifact is ai-agents-ent-product-dev-spec.md (~50KB, 15 sections), a vendor-neutral specification covering orchestration protocols, a hierarchical work-item graph (Initiative→Epic→Story→Ticket→Task), a Fluid Dependency Graph, immutable event ledger, governance/audit mechanisms, and scalability from single-agent to committee-based orchestration for millions of concurrent agents. A companion roo-code-setup/ directory provides a shell script (enterprise_roo_setup.sh) that generates a production-ready Roo Code environment with 12 specialized enterprise modes (orchestrator, architect, code, debug, devops, security, performance, product, compliance, research, docs, data) as roomodes.yaml for VS Code's Roo Code extension. The spec defines three architectural abstractions absent from most peers: (1) a Fluid Dependency Graph with typed edges (Definition, Implementation, Testing, Production) that dynamically unblocks tasks; (2) a Delivery/Dependency Agent as a distinct meta-agent owning the FDG and AI-generated dashboards; (3) Committee-based orchestration using consensus protocols (Paxos/Raft/PBFT) for high-availability at scale. Human oversight is structured as push/pull: the Human Overseer receives OversightEscalation events automatically and can pause/resume any flow. The framework is primarily a methodology/spec document with a secondary Roo Code implementation layer.

01

Overview

enterprise-ai-agents-spec — Overview

Origin

Created by Zohar Babin (GitHub: zoharbabin). Repository: zoharbabin/enterprise-ai-agents-spec. MIT license. 51 stars. Created 2024-12-27, last pushed 2025-07-19. Shell + Markdown.

Philosophy

"A human-conducted orchestra of AI Agents, seamlessly delivering enterprise-grade products. A symphony of intelligence, precision, and innovation."

The framework is explicitly a proposed specification and practical blueprint — not a runnable library. It distills practices from human software teams (Agile, Scrum, Kanban, Waterfall) into protocols that an AI swarm can implement. The central commitment is minimal human oversight with maximum auditability: agents handle all day-to-day tasks autonomously while every action is appended to an immutable ledger, and humans receive automatic escalations for high-impact decisions.

The spec targets two audiences equally: human practitioners who need an auditable process, and AI agents that will parse the guidelines and self-implement.

Scope

The document covers the meta-layer of orchestration and collaboration, explicitly not the individual agent implementations. The philosophy is: "Numerous implementations and toolkits already exist for each Agent role... this document focuses primarily on the meta-layer."

Scale Vision

The spec explicitly designs for swarms of millions of concurrent agents through committee-based orchestration (Section 11). The progression is: single-orchestrator for small projects → committee (multiple Orchestrator nodes with consensus protocol) for mission-critical environments.

Version

Analyzed from main branch (no semver on the spec). roo-code-setup/templates/version.json tracks mode versions. Last pushed 2025-07-19.

02

Architecture

enterprise-ai-agents-spec — Architecture

Distribution

  • Type: Methodology document + Roo Code plugin generator
  • Primary artifact: ai-agents-ent-product-dev-spec.md (vendor-neutral spec, ~50KB)
  • Secondary artifact: roo-code-setup/ with enterprise_roo_setup.sh bash script
  • Languages: Shell (setup script), Markdown (spec + instruction templates)

Installation

# One-command setup via GitHub Pages
curl -sSL https://zoharbabin.github.io/enterprise-ai-agents-spec/enterprise_roo_setup.sh | bash -s -- ./my-project

# Selective mode install
curl -sSL .../enterprise_roo_setup.sh | bash -s -- --modes orchestrator,architect,code,security ./my-project

# Interactive selection
curl -sSL .../enterprise_roo_setup.sh | bash -s -- --interactive ./my-project

The setup script generates a roomodes.yaml file in the target project directory, compatible with the Roo Code VS Code extension.

Directory Tree

zoharbabin/enterprise-ai-agents-spec/
├── ai-agents-ent-product-dev-spec.md   # Primary spec (~50KB, 15 sections)
├── CHANGELOG.md
├── README.md
├── .github/                            # GitHub Actions for spec change tracking
└── roo-code-setup/
    ├── enterprise_roo_setup.sh         # Bash setup script (~26KB)
    ├── IMPLEMENTATION-GUIDE.md         # Enhanced ecosystem guide
    ├── enterprise-mode-development-guide.md  # Mode authoring guide
    ├── README.md                        # Quick-start documentation
    ├── templates/
    │   ├── roomodes.yaml               # Roo Code mode definitions template
    │   └── version.json                # Mode version tracking
    └── instruction-templates/          # 17 instruction template files
        ├── orchestrator.md             # Orchestrator mode instructions v5.0
        ├── architect.md
        ├── code.md
        ├── debug.md
        ├── devops.md
        ├── security.md
        ├── performance.md
        ├── product.md
        ├── compliance.md
        ├── research.md
        ├── docs.md
        ├── data.md
        ├── quality-gates-framework.md   # Shared 4-phase validation framework
        ├── mcp-integration-guide.md     # MCP research integration guide
        ├── enterprise-compliance-framework.md
        └── README.md

Required Runtime

  • Bash (for setup script)
  • Roo Code VS Code extension (for mode execution)
  • No npm/Node.js required for setup; Roo Code requires VS Code

Target AI Tools

  • Primary: Roo Code (VS Code extension)
  • Secondary: Any tool capable of reading the spec markdown (vendor-neutral methodology layer)

Architecture Layers

  1. Spec Layer (ai-agents-ent-product-dev-spec.md): Vendor-neutral protocols, message formats, governance rules, orchestration models
  2. Mode Layer (instruction-templates/*.md): Roo Code-specific role instructions implementing the spec
  3. Config Layer (templates/roomodes.yaml): Roo Code mode configuration (name, instructions, tools, file restrictions)
  4. Setup Layer (enterprise_roo_setup.sh): Generates project-specific configuration from templates
03

Components

enterprise-ai-agents-spec — Components

Spec Document Sections (15 + appendix)

Section Title
1 Introduction
2 Terminology & Definitions
3 High-Level Lessons from Human Methodologies
4 Work-Item Graph & Cross-Level Dependencies
5 Human Feedback & Critical Timing
6 Agent Role Specialization & Communication
7 Governance & Audit Processes
8 Architecture & Domains
9 AI-Generated Dashboards
10 Parallelization & Cost-Optimization
11 Orchestrator & Coordination Models
12 Distributed Collaboration & Integrations
13 Peer Review Culture & Meritocracy
14 Putting It All Together: Simplified Example Workflow
15 Conclusion
16 Further References and Implementations for Specific Agent Roles

Core Abstractions (Spec Layer)

Agent

AI component specialized in a role. Receives Tasks, produces artifacts, communicates via event-based or API-driven methods. Roles include: Product/Market Analyst, Designer (UX/UI), Developer, QA/Tester, Security Auditor, Compliance Specialist, FinOps Analyst, Delivery/Dependency Manager.

Orchestrator

Meta-agent coordinating the swarm. Responsibilities: task distribution, concurrency management, output merging, human interface for approvals/pausing. Begins as single node, scales to committee.

Delivery/Dependency Agent

Distinct from the Orchestrator. Owns the Fluid Dependency Graph and AI-Generated Dashboards. Detects dependency conflicts, resource spikes, and suspicious activity; triggers OversightEscalation events.

Work-Item Graph

Hierarchical: Initiative → Epic → Story → Ticket → Task. Cross-level dependencies allowed. Each work item captures a distinct unit of functionality.

Fluid Dependency Graph (FDG)

Dynamic graph: G = (W, E) where W = work items, E = typed directed edges. Predefined edge types: Definition (δ), Implementation (ι), Testing (τ), Production (ρ). Custom types (Cost, Security) addable dynamically. All changes immutably logged.

Message Queue

Fault-tolerant event backbone. Supports dynamic event types, priority scheduling, and immutable logging. Message structure: (id, timestamp, source, target, payload, priority, schema_version, validation_function). Routing includes failover: active agent → backup agent → Orchestrator.

Key event types: TaskCreated, SecurityAlert, DependencyResolved, OversightEscalation, PauseFlow, SuspiciousActivity, WorkItemStatusChanged.

Governance & Audit

Immutable append-only ledger for all actions. Policy enforcement is configurable as blocking or non-blocking. Manual gating for production releases. Compliance frameworks: SOX, HIPAA, PCI-DSS, GDPR, FedRAMP.

Human Overseer

Human with pause/resume authority. Receives OversightEscalation events (push) and can query dashboards (pull). All Overseer decisions immutably recorded. Trust-but-verify: system operates autonomously; Overseer intervenes when needed.

Roo Code Modes (12)

Mode Role
orchestrator Enterprise project coordination, workflow management
architect System architecture, technology evaluation
code Senior development, testing integration
debug QA, comprehensive testing
devops Infrastructure, deployment automation
security Cybersecurity, threat assessment, compliance
performance Performance engineering, monitoring
product Product management, stakeholder coordination
compliance Regulatory compliance, audit preparation
research Market and technical research with MCP
docs Technical documentation generation
data Data engineering, analytics pipelines

Shared Framework Documents (Roo Code Layer)

Document Purpose
quality-gates-framework.md 4-phase validation system (requirements, design, implementation, deployment)
mcp-integration-guide.md Google Researcher MCP server integration, research decision matrix
enterprise-compliance-framework.md Regulatory requirements, audit trail procedures

Setup Script Capabilities

enterprise_roo_setup.sh (25KB):

  • Installs all 12 modes or a selected subset (--modes, --interactive, --list)
  • Generates roomodes.yaml with full mode definitions
  • Creates project directory structure (security/, compliance/, docs/ subdirectories)
  • Supports offline mode (--local) using bundled templates
  • Fetches latest templates from GitHub Pages to avoid cache issues

GitHub Integration

.github/ directory contains GitHub Actions for spec change tracking — CI runs on spec document changes to flag potentially breaking modifications.

05

Prompts

enterprise-ai-agents-spec — Prompts & Instructions

Instruction Template Structure

Each Roo Code mode has a corresponding instruction-templates/<mode>.md file. The Orchestrator mode is versioned as v5.0; the enhanced enterprise modes (Product, Architect, Security, Research) as v6.0. Instructions are injected as the system prompt for the corresponding Roo Code mode.

Orchestrator Mode Instructions (v5.0)

The Orchestrator prompt establishes two primary functions:

1. Workflow Orchestration

The Orchestrator uses automatic switch_mode(target, instruction) calls to delegate to specialist modes. The delegation decision matrix defines:

Task Complexity Assessment:
  Low: Handle directly (simple configs, status updates, coordination)
  Medium: Evaluate context and timeline requirements
  High: Delegate to specialized mode

When to Delegate:
  - Complex technical implementation requiring deep domain expertise
  - Multi-step processes benefiting from specialized tooling
  - Quality-critical work needing focused validation
  - Stakeholder-facing deliverables requiring specialized communication
  - Compliance or security tasks requiring audit trails

Handle Directly:
  - High-level coordination and workflow management
  - Simple configuration or setup tasks
  - Status updates and progress reporting
  - Quality gate validation and approval workflows
  - Emergency response coordination and escalation

2. Quality Gate Implementation

The Orchestrator validates gate completion using ask_followup_question with structured suggestions:

[Phase] completed with [achievements]. [Validation summary]. Proceed to [next phase]?
→ suggest: [Primary action with mode switch]
→ suggest: [Alternative action without mode switch]
→ suggest: [Refinement option]
→ suggest: [Stakeholder engagement option]

Shared Framework Documents

Three shared framework documents are referenced across all mode instructions:

Quality Gates Framework

Four-phase universal validation pattern. Each mode customizes for its domain:

  • Product Mode: business value validation + stakeholder alignment
  • Architect Mode: technical feasibility + enterprise standards
  • Security Mode: threat validation + compliance verification
  • Research Mode: research accuracy + strategic impact

MCP Integration Guide

Research decision matrix:

  • Direct research in current mode: tactical needs <5 min, specific technical questions, quick validation
  • Delegate to Research Mode: strategic analysis >5 min, competitive intelligence, multi-source synthesis

Enterprise Compliance Framework

Cross-mode compliance integration for regulatory requirements (SOX, HIPAA, PCI-DSS, GDPR, FedRAMP). Defines audit trail procedures and validation processes.

Spec Document as Prompt Source

The ai-agents-ent-product-dev-spec.md itself is designed to be consumed by AI agents:

"This document is intended for: AI Agents that will parse these guidelines, then implement or prototype the system."

Section 2 (Terminology & Definitions) uses formal mathematical notation (set theory, graph notation, routing functions) intended for precision when AI agents interpret the spec. For example:

  • Fluid Dependency Graph formally defined as G = (W, E)
  • Message routing formalized as a piecewise function
  • Cost constraint expressed as Σ C(w) ≤ B

This dual-audience design (human practitioners + AI agents) makes the spec unique: it is simultaneously documentation and agent instruction.

Mode Development Guide

enterprise-mode-development-guide.md provides a template for custom mode instructions:

## Mode Purpose Statement
**Primary Function:** [Single sentence describing main responsibility]
**Domain Expertise:** [Specific area of specialization]
**Target Users:** [Who will primarily use this mode]
**Business Value:** [Contribution to organizational objectives]

Scope boundaries are defined as YAML blocks listing included responsibilities, excluded responsibilities, and collaboration points with other modes.

Event Protocol Instructions

Agents are instructed to use specific event names from the Message Queue schema:

  • OversightEscalation — for issues requiring Human Overseer attention
  • SuspiciousActivity — for agent misbehavior detection
  • PauseFlow — for halting task assignment
  • TaskCreated, SecurityAlert, DependencyResolved — standard workflow events

These event names are defined in Section 2 of the spec and referenced consistently across all mode instructions.

09

Uniqueness

enterprise-ai-agents-spec — Uniqueness

Distinguishing Characteristics

1. Fluid Dependency Graph with Typed Edges

The FDG is the most formally specified data structure in this batch. Unlike task DAGs in other frameworks (which track dependencies for scheduling), the FDG uses named edge types (Definition, Implementation, Testing, Production, plus custom types). This supports semantically meaningful queries: "what tasks depend on this Definition?" vs. "what tasks depend on this Production deployment?" The FDG is designed to be dynamically extended (new edge types added at runtime) without breaking existing nodes. No other framework in this batch formalizes dependency edges to this degree.

2. Delivery/Dependency Agent as a First-Class Role

Most frameworks give the Orchestrator both scheduling and dependency-tracking responsibilities. Enterprise-ai-agents-spec explicitly separates these: the Delivery/Dependency Agent owns the FDG, project analytics, and dashboard generation; the Orchestrator handles task assignment and policy enforcement. This creates a cleaner separation of concerns and allows domain-specific Delivery Agents to operate per microservice domain while feeding a global Orchestrator.

3. Committee-Based Orchestration with Consensus Protocols

The spec explicitly names Paxos, Raft, and PBFT as candidate consensus mechanisms for multi-Orchestrator coordination. This is the only framework in this batch that addresses the distributed systems implications of multi-agent orchestration at scale. Committee leadership rotation and immutable ledger replication are defined as explicit design requirements.

4. Dual-Audience Spec: Human + AI-Parseable

The primary artifact is explicitly written for two audiences: human practitioners and AI agents. Mathematical notation (set theory, graph theory, routing functions) is used for precision in AI-parseable definitions. Section 2 is essentially a formal specification that an AI agent can implement from directly. No other framework in this batch publishes a spec document in this dual-audience format.

5. AI-Generated Dynamic Dashboards

The Delivery Agent does not serve pre-defined reports. It dynamically decides what data to surface per stakeholder and project state. Natural language queries replace pre-defined dashboard configurations. Push alerts via email/WhatsApp are specified (not just in-tool notifications). This is described as a novel capability enabled by AI: "There is no need to pre-define reports or manage dashboards anymore."

6. Explicitly Scalable to Millions of Agents

The spec dedicates sections 11 and 12 to scalability concerns specific to massive swarms. It provides concrete mechanisms: committee-based orchestration prevents single points of failure, load balancing distributes analytics processing across multiple Delivery Agents, and consensus protocols maintain consistency. No other framework in this batch addresses this scale or specifies the systems-level mechanisms required for it.

7. Rogue Agent Handling with Detection Protocol

The spec defines a specific event type (SuspiciousActivity) and escalation workflow for detecting and containing misbehaving agents. The detection signals (rapid contradictory commits, repeated error loops, event frequency spikes) and mitigation actions (quarantine, PauseFlow, Overseer notification) are documented explicitly. This treats agent reliability as a first-class architectural concern.

Compared to Seeds

  • vs. BMAD-METHOD: BMAD uses persona-md agents; enterprise-ai-agents-spec defines protocols for agent coordination, not the agents themselves. BMAD has no FDG, no committee orchestration, no formal spec.
  • vs. open-multi-agent: open-multi-agent implements a programmatic DAG coordinator; enterprise-ai-agents-spec describes a higher-level methodology with typed dependency edges, governance, and audit protocols.
  • vs. devteam-michael-harris: DevTeam provides a runnable Claude Code plugin; this spec provides a vendor-neutral blueprint implemented via Roo Code. DevTeam's Bug Council is analogous to the spec's escalation chain but implemented; the spec's governance layer has no direct DevTeam equivalent.

Weakest Points

  • Roo Code implementation is single-mode-at-a-time: the Roo Code switch_mode() pattern is sequential, not the parallel swarm the spec describes.
  • No state management implementation: the spec mandates an immutable ledger and FDG but provides no implementation. Implementations must build these from scratch.
  • 51 stars suggests limited adoption: the gap between the spec's ambitious scope and actual implementations remains large.
  • AI-generated dashboards are aspirational: the spec describes dashboard generation as a capability but the Roo Code setup provides no dashboard implementation.
04

Workflow

enterprise-ai-agents-spec — Workflow

Main Development Workflow (Orchestrator Mode)

The Orchestrator mode in Roo Code implements the spec's coordination model:

1. Requirements & Product Strategy
   → switch_mode("product") → requirements definition
   → switch_mode("research") → market/technical feasibility
   → Quality Gate: Requirements Sign-off

2. Architecture & Security Design
   → switch_mode("architect") → enterprise architecture
   → switch_mode("security") → threat modeling
   → switch_mode("compliance") → regulatory validation
   → Quality Gate: Architecture Approval

3. Implementation & Quality
   → switch_mode("code") → implementation
   → switch_mode("debug") → testing and QA
   → switch_mode("performance") → performance testing
   → Quality Gate: Implementation Validation

4. Infrastructure & Deployment
   → switch_mode("devops") → CI/CD setup, deployment
   → switch_mode("data") → data pipelines
   → Quality Gate: Production Readiness

5. Documentation & Compliance
   → switch_mode("docs") → documentation generation
   → switch_mode("compliance") → final audit preparation

Quality Gate Pattern (4-Phase)

Each quality gate uses a standardized approval workflow:

Planning & Requirements Validation
Design & Architecture Approval
Implementation & Testing Validation
Deployment & Success Confirmation

Gate implementation uses Roo Code's ask_followup_question with suggested next steps:

  • Primary action with mode switch
  • Alternative action without mode switch
  • Refinement option
  • Stakeholder engagement option

Task Parallelization

The Delivery/Dependency Agent continuously monitors the FDG:

  • When work item W_i completes, check FDG for newly unblocked items
  • Assign newly available items to idle agents
  • Budget constraint: Σ C(w) for active work ≤ B (configurable budget limit)

Cost factors tracked: compute resources, API calls, licensing/SaaS, human reviewer time, security overhead.

Escalation Workflow

Agent detects anomaly
  → Local resolution attempt (Delivery Agent)
      → Resolved: continue
      → Unresolved: SuspiciousActivity event → Orchestrator
          → Committee consensus (if multi-orchestrator)
          → PauseFlow issued to affected agents
          → OversightEscalation → Human Overseer
              → Overseer: Approve / Reject / Query
              → Decision immutably logged

Human Overseer oversight is both push (automatic alerts for SecurityAlert, budget overrun, OversightEscalation) and pull (dashboard queries on demand).

Spec Workflow (Vendor-Neutral)

Per the main spec document:

1. Orchestrator receives project goal
2. Decompose into Work Items (Initiative → Epic → Story → Ticket → Task)
3. Build Fluid Dependency Graph
4. Delivery Agent tracks FDG, unblocks tasks as dependencies resolve
5. Agents receive TaskCreated events from Message Queue
6. Agents execute, append results to immutable ledger
7. QA/Security/Compliance checks run continuously
8. OversightEscalation events route to Human Overseer
9. Human Overseer: pause/resume/approve at critical gates
10. Sprint/milestone checkpoint reviews compile governance reports

Conflict Resolution

Minor conflicts are resolved locally before escalating:

  1. Delivery Agent or Requirement Clarification Agent negotiates
  2. Two Developer Agents disagreeing → run benchmark, choose winner
  3. Complementary solutions → Merger Agent synthesizes without human input
  4. Only if local resolution fails or deadlines exceeded → escalate to Orchestrator
  5. Only if Orchestrator cannot resolve → Human Overseer

Approval Gates

Gate Trigger Mechanism
Sprint/milestone review Scheduled Compliance/governance report compiled by Delivery Agent
Production release Per deployment Explicit Orchestrator or Human Overseer sign-off
Budget overrun Cost C(w) > BUDGET_LIMIT OversightEscalation (blocking)
Security alert Vulnerability detected OversightEscalation (blocking or non-blocking by config)
Oversight required Any critical risk Human Overseer pause/inspect
06

Memory Context

enterprise-ai-agents-spec — Memory & Context

Spec-Level Memory Model

The spec defines memory through two mechanisms: the immutable event ledger and the Fluid Dependency Graph.

Immutable Event Ledger

The central persistence mechanism. Defined formally as: L = { (t₁, A₁, action₁, result₁), (t₂, A₂, action₂, result₂), ... }

Properties:

  • Append-only: No data is deleted; corrections are new entries
  • Decision tracking: All critical decisions, especially human ones, logged with rationale
  • Replicated: In committee mode, the ledger is replicated across Orchestrator nodes for fault tolerance

Every significant action is appended: task state changes, escalation events, human approvals, security alerts, compliance checks.

Fluid Dependency Graph (FDG)

Maintained by the Delivery/Dependency Agent. Dynamic state: as work items complete, edges are resolved and newly unblocked items become available. FDG changes are themselves logged to the immutable ledger.

The FDG is the system's "working memory" for project state — what's blocked, what's available, what's complete, and why.

Roo Code Implementation Layer

The Roo Code instruction templates do not define a specific in-session memory mechanism beyond what Roo Code provides natively. Context is passed between mode switches via the switch_mode(target, instruction) call where the instruction string provides handover context.

Context Preservation Protocol

The IMPLEMENTATION-GUIDE.md describes "context preservation" as a feature of the coordination system:

"Advanced Coordination System: Automatic workflow management with context preservation and seamless mode transitions"

In practice, this means the Orchestrator mode passes sufficient context in each switch_mode call to allow the target mode to continue without requiring the full conversation history. The Orchestrator instruction template explicitly handles "emergency response coordination" and "status updates" directly rather than delegating, preserving coordination context.

Research Mode Memory

The Research mode (v5.0) integrates MCP for persistent research capabilities. The MCP Integration Guide defines a decision matrix for when to use direct research (within current context) vs. delegating to Research Mode for deep investigation. Research findings are intended to be written to project files for persistence beyond the session.

Checkpoint Reviews as Memory

Section 5 of the spec defines Scheduled Review Gates at sprints or milestones. These compile compliance/governance reports from the ledger, effectively creating periodic memory snapshots of project state. Ad hoc oversight queries also serve as pull-based memory access against the ledger.

Context Across Scales

The spec's scalability design has implicit context management implications:

  • Single-orchestrator: full project context in one node
  • Committee: context is replicated across Orchestrator nodes; Delivery Agents maintain domain-local context; the global Orchestrator maintains cross-domain context

The separation of Delivery/Dependency Agent (domain context) from Orchestrator (global coordination context) is an explicit context partitioning strategy.

No Session-Level Memory Specification

The spec does not prescribe specific mechanisms for within-session context management (no explicit summarization strategy, context window management, or cross-agent state synchronization protocol). These are implementation details left to the runtime (Roo Code or other tools). The spec focuses on durable persistence (ledger, FDG) rather than ephemeral context window management.

07

Orchestration

enterprise-ai-agents-spec — Orchestration

Orchestration Models

Single-Orchestrator (Single-Cell)

For smaller or early-stage projects:

  • One meta-agent manages task assignments, message routing, human interface
  • Advantages: low overhead, easy configuration
  • Drawbacks: single point of failure, limited scalability

Committee (Multi-Orchestrator)

For large, mission-critical environments:

  • Multiple Orchestrator nodes using consensus protocol (Paxos, Raft, or PBFT)
  • Shared state replication across nodes
  • Committee leadership rotation if a node fails or becomes unreliable
  • Higher complexity and resource usage; required for resilience against rogue agents

The spec explicitly targets scalability "to millions of agents" through committee-based orchestration.

Orchestrator + Delivery Agent Separation

A key architectural distinction: the Orchestrator and Delivery/Dependency Agent have separate domains.

Orchestrator responsibilities:

  • Schedule tasks across agents
  • Enforce gating rules
  • Receive escalations from Delivery Agents
  • In committee mode: replicate all updates across nodes
  • Pause or isolate agents when thresholds/policy checks fail

Delivery/Dependency Agent responsibilities:

  • Maintain the Fluid Dependency Graph
  • Detect task completions and unblock downstream work items
  • Own project analytics and AI-generated dashboards
  • Detect anomalies (repeated errors, cost spikes, suspicious activity)
  • Issue OversightEscalation when local resolution fails
  • Serve as "source of truth" for domain-level metrics

In multi-domain projects, one Delivery Agent per domain feeds into the Orchestrator (or committee).

Agent Role Specialization

Agents communicate via the Message Queue (event-driven). The spec defines 8 core roles:

  • Product/Market Analyst
  • Designer (UX/UI)
  • Developer
  • QA/Tester
  • Security Auditor
  • Compliance/Regulatory Specialist
  • FinOps/Cost Analyst
  • Delivery/Dependency Manager

Custom roles (e.g., Observability Agent) can be added without structural changes: define role, add/reuse event types, specify agent tasks and integration points. No core components require modification.

Task Assignment Flow

Message Queue → Agent receives TaskCreated event
Agent executes → appends result to immutable ledger
Delivery Agent monitors FDG for unblocked work items
Orchestrator schedules newly available items

If an agent fails, tasks revert to Available status. Pausing is enforced by the Orchestrator via PauseFlow event.

Roo Code Orchestration (Implementation Layer)

In the Roo Code implementation, orchestration uses switch_mode() calls rather than a separate agent process:

Orchestrator mode → switch_mode("product", "Define requirements for...")
Product mode executes → returns to Orchestrator
Orchestrator mode → switch_mode("architect", "Create architecture based on...")
...

This is sequential mode-switching within a single VS Code conversation, not parallel agent execution. The spec's parallel swarm model is a methodology target; the Roo Code layer is a practical single-session implementation.

Governance Enforcement in Orchestration

The Orchestrator enforces governance at the meta-level:

  • Automated checks run at both Delivery Agent (domain-level thresholds) and Orchestrator (global/cross-domain) levels
  • OversightEscalation events are generated when: cost exceeds budget, security alerts are raised, tasks exceed expected duration, suspicious activity is detected
  • Manual gating: certain flows (production releases) require explicit sign-off before proceeding

Peer Review & Meritocracy

Section 13 of the spec defines a peer review culture:

  • Agents can review each other's work (automated + specialized review agents)
  • Complementary solutions are merged by a "Merger Agent"
  • Two conflicting implementations: run benchmark, choose winner by outcome
  • Meritocracy: best solution wins regardless of which agent produced it

Rogue Agent Handling

The committee architecture provides defense against rogue agents:

  1. Delivery Agent detects anomaly (contradictory commits, repeated error loops)
  2. Local resolution attempted
  3. SuspiciousActivity event → Orchestrator
  4. Committee validates through consensus
  5. Orchestrator issues PauseFlow, quarantines the agent
  6. OversightEscalation → Human Overseer for final decision
08

Ui Cli Surface

enterprise-ai-agents-spec — UI & CLI Surface

Setup Script CLI

The primary CLI surface is enterprise_roo_setup.sh:

# Install all 12 modes (default)
./enterprise_roo_setup.sh ./my-project

# Install specific modes
./enterprise_roo_setup.sh --modes orchestrator,architect,code,security ./my-project

# Interactive mode selection
./enterprise_roo_setup.sh --interactive ./my-project

# Use local templates only (offline)
./enterprise_roo_setup.sh --local ./my-project

# List available modes
./enterprise_roo_setup.sh --list

Output: generates roomodes.yaml in the target project directory, compatible with Roo Code.

Roo Code VS Code Interface

After setup, the UI surface is entirely within VS Code via the Roo Code extension:

  • Mode selection via Roo Code's mode switcher UI
  • Conversation-driven interaction within VS Code panel
  • switch_mode() calls are transparent to the user — mode transitions appear as context switches in the conversation
  • No separate terminal or TUI

AI-Generated Dashboards

Section 9 of the spec defines a novel UI concept: the Delivery Manager Agent generates dynamic dashboards on demand rather than pre-defined reports.

"There is no need to pre-define reports or manage dashboards anymore, and natural language can be used to ask for project status."

The Delivery Agent decides which data to surface based on stakeholder role and current project state. Real-time alerts can be pushed via email or WhatsApp (configurable). Dashboard updates fire on every work item state change.

Types of dashboards:

  • Security: vulnerabilities, patch status
  • FinOps: cost usage, forecast, ROI
  • Project status: task board (To Do → In Progress → Review → Done)
  • Compliance reports: scheduled at sprints/milestones

Human Overseer Interface

The Human Overseer interacts through two channels defined in the spec:

  • Pull: queries dashboards/logs on demand
  • Push: automatic OversightEscalation alerts for critical issues

The pausing mechanism:

  • Overseer decides to pause → Orchestrator issues PauseFlow event
  • All tasks become "paused"
  • No new tasks assigned until resume event
  • Resume restores normal operation

No Web UI

The framework does not ship a web UI. The spec describes AI-generated dashboards conceptually; the Roo Code implementation layer relies on VS Code's built-in interface. No separate dashboard application is provided.

Spec as Documentation UI

The README.md (zoharbabin/enterprise-ai-agents-spec) and CHANGELOG.md serve as the human-facing documentation layer. GitHub Pages hosts the latest setup script for direct curl execution, which is the primary distribution mechanism for end users.

Mode Development Interface

The enterprise-mode-development-guide.md provides a template-based interface for creating custom modes:

  • Purpose Statement Template (markdown)
  • Scope Boundaries Framework (YAML)
  • Mode Classification System
  • Quality Gates Implementation Pattern
  • Inter-Mode Collaboration Protocol

This is the "extensibility UI" — how practitioners add new agent roles to the system without modifying core components.

Related frameworks

same archetype · same primary tool · same memory type

BMAD-METHOD ★ 48k

Provides a full agile delivery lifecycle with named expert-persona AI collaborators that elicit the human's best thinking rather…

Agent OS ★ 4.6k

Extracts implicit codebase conventions into token-efficient markdown standards files and injects them selectively into AI agent…

Claude Conductor ★ 367

Gives Claude Code a persistent, cross-linked, auto-analyzed documentation system so it retains codebase context across sessions.

Spec-Driver (Greenfield Spec-Driven Development) ★ 25

Prevents spec rot in AI-assisted development by making implementation changes flow back into evergreen, authoritative specs via…

Anthropic Knowledge Work Plugins ★ 16k

Role-specialized plugin bundles with live MCP connectors that turn Claude into a domain expert for enterprise knowledge workers.

Codex Integration for Claude Code (skill-codex) ★ 1.3k

Single Claude Code skill that handles Codex CLI invocation correctly (stdin blocking, thinking token suppression, session resume)…