Triagent

AI Assistant for Every Persona in Your Team

Powered by Claude Agent SDK Anthropic

Context-aware AI agent platform that understands your team, integrates with your tools, and automates tasks autonomously

Why Current AI Tools Fail in Enterprise

The Gap Between Chat Assistants and Productive AI Agents

✗

Generic AI Lacks Team Context

No understanding of your ADO structure, coding standards, team workflows, or internal processes

✗

Chat-Only = No Action Execution

AI provides suggestions but cannot execute - users must copy/paste and run commands manually

✗

No Compute Sandbox for Code

Cannot run scripts, execute commands, or interact with file systems in a secure environment

✗

No Cross-System Integration

Limited to search - cannot write to systems, execute authenticated API calls, or correlate data

✗

No Persistent Memory

Every session starts fresh - no project context, no learnings from previous interactions

Triagent: Context-Aware AI Agent Platform

Team Context

Loads team-specific skills and context, understands ADO projects, repositories, and workflows

Tool Integration

Connects with ADO, Azure, Databricks, Kusto via MCP protocol - read AND write access

Task Automation

Creates PRs, updates work items, runs CLI commands - all autonomously with user controls

Built on Claude Agent SDK — the same infrastructure powering Anthropic's official Claude Code CLI

Architecture Overview

Triagent CLI
Local Machine
Triagent CLI
Terminal Interface
↓
Triagent Core
Skills + Personas
↓
Claude Agent SDK
+ Claude Code CLI
Git Azure CLI Python .NET Bash
Developer tools installed locally with full filesystem access
Triagent Web
Azure Cloud
App Service
Triagent Web (Chainlit)
Browser UI
Container Apps Session Pool
Sandboxed Container (per session)
Triagent Core
Skills + Personas
Claude Agent SDK
+ Claude Code CLI
Git Azure CLI Python .NET Bash
↓
Shared Services
Azure AI Foundry
Claude Models (Opus / Sonnet)
MCP Tools
ADO, Azure, Kusto, Databricks, Git...

Security Architecture

Enterprise-Grade Security Controls

Authentication

  • Stage 1: Azure AD OAuth (MSAL) for Web UI access
  • Stage 2: Device Code Flow for Azure CLI in container
  • User identity propagation (no shared service accounts)
  • RBAC for Azure resource access control

Network Security

  • Container Apps in dedicated subnet
  • Private endpoints for Azure services
  • NSG rules for traffic control
  • Azure AD protection on public endpoints
Roadmap: Full VNET integration with private network deployment

Data Protection

  • Session Isolation: Each session = isolated container
  • Ephemeral Filesystem: Destroyed on session end
  • Storage Encryption: TDE for SQL, encrypted Redis
  • No PII retention in AI model context

Secrets & Identity

  • Azure Key Vault for API keys
  • Managed Identity for service-to-service auth
  • No hardcoded credentials in containers
  • Audit logging via Application Insights

Data Flow & Integration

End-to-End Request Flow

User
Browser
→
Chainlit
App Service
→
Session Pool
Container Apps
→
Claude API
AI Foundry
↓ Tool Calls ↓
Azure DevOps
Work Items, PRs
Azure CLI
Resource Mgmt
Kusto
Telemetry
Git
Repos
Data Storage
Azure SQL
Chat History
Redis Cache
Session State
App Insights
Logs & Telemetry
Container FS
Ephemeral (per-session)

Technology Stack

Powered by Microsoft Azure

Microsoft Azure
AI Layer
Azure AI Foundry

Claude Opus 4 / Sonnet 4 (via Anthropic on Azure)

Compute Layer
Container Apps

Session Pool | 2 vCPU / 4GB per session

Triagent-Core Claude-SDK Claude-Code Git Azure-CLI Python .NET Bash
App Service

Chainlit Web UI | P1v3 SKU

Storage & Observability
ACR

Standard SKU

Redis

Basic C0

SQL Database

Basic 5 DTU

App Insights

Telemetry

Identity

Managed ID

AI Assistant Comparison

Internal Tools (GPT-4o) vs Triagent (Claude)

Capability Sidekick / Pal (GPT-4o) Triagent (Claude)
Team Context ✗ Generic AI assistant - no team knowledge ✓ Loads team-specific skills, ADO context
Autonomous Agent ✗ Chat only - user must execute ✓ Creates PRs, updates work items autonomously
Compute Sandbox ✗ No sandbox - text generation only ✓ Isolated container with dev tools
Multi-Agent / Skills ✗ Single model conversation ✓ PR-reviewer, code-analyzer, subagents
Tool Integration ● Web & Enterprise Search (read-only) ✓ MCP: ADO, Azure, Kusto (read + write)
Deep Research ● Document search only ✓ Cross-system: Git + Logs + Wiki + APIs
Context & Memory ● Basic chat history ✓ Project context, file awareness, resume

Claude Agent SDK: Key Features

Programmatic access to the same infrastructure powering Claude Code

1. Autonomous Tool Execution

Built-in tools for file operations, bash commands, git, web search - no need to implement tool loops yourself

2. Specialized Subagents

Spawn focused agents for subtasks - code-reviewer, documentation, testing - enabling parallel work distribution

3. MCP Integration

Connect to 100+ external systems - Jira, Figma, Slack, Databricks, databases - without custom implementation

4. Enterprise Security

Pre/post-tool hooks, permission controls, allowlist tools, audit trails - enterprise-grade compliance

5. Persistent Sessions

Maintain context across exchanges - resume sessions, fork explorations - critical for multi-stage workflows

Dynamic Skills System

Skills as Markdown Files

Each skill is a markdown file with metadata - easily editable, version controlled, and extensible

Version Controlled: Skills live in Git, enabling collaboration and change tracking
pr_review.md
finops.md
telemetry.md
lsi_creation.md
code_review.md
backlog.md

Session Initialization

Skills loaded dynamically based on user's persona

User selects: Developer
↓ Skills Loaded:
• pr_review
• code_review
• telemetry
• pipeline_investigation

Live Demo

See Triagent in Action

Switch to Triagent Web UI

Developer PR Review
Telemetry Investigation
Work Item Creation

Use Cases by Persona

One platform, tailored capabilities for every role

Developer

  • Automated PR reviews
  • Build failure investigation
  • Telemetry analysis
  • Create PRs with fixes

Support

  • Telemetry investigation
  • LSI ticket creation
  • Root cause analysis
  • Backlog recommendations

BA / PO

  • AI work item creation
  • Acceptance criteria
  • Test specifications
  • Backlog prioritization

FinOps

  • Azure cost insights
  • Daily/weekly reports
  • Resource optimization
  • Budget tracking

Scalability & Operations

Enterprise-Ready Infrastructure

Session Pool Scaling

  • Container startup < 1 second (warm pool)
  • Session timeout 30 min idle
  • Max session duration 2 hours
  • Scale trigger CPU > 70%

Resource Limits (per session)

  • CPU 2 vCPU
  • Memory 4 GB
  • Ephemeral storage 8 GB
  • Network Egress allowed

Monitoring & Observability

  • Telemetry Application Insights
  • Custom metrics Session duration, tokens
  • Alerting Error rate > 5%
  • Log retention 30 days

Operational Excellence

  • Deployments Zero-downtime rolling
  • Container registry ACR with geo-replication
  • Health probes Liveness + Readiness
  • IaC Bicep templates

Vision: Deloitte AI Agent Marketplace

From Single Team Tool to Enterprise Platform

Today: v1.0

Triagent for Omnia Data Team

  • Developer PR Reviews
  • Support Telemetry Investigation
  • Product Owner Work Items
  • FinOps Cost Analysis

Tomorrow: Marketplace

Deloitte-wide Skills & Agents

ServiceNow Agent Time Management KX Articles SAP Integration Salesforce Agent HR Self-Service

Any team can publish skills → Any team can consume

Next Steps

1

Architecture Approval

Review security controls, data flow, and infrastructure design

2

Claude Model Access

Get Claude model usage approval for Audit Cortex team via AI Foundry

3

Pilot Deployment

Deploy to sandbox environment with Omnia Data team

4

Production Rollout

VNET integration, private endpoints, full compliance review

Questions?

Ready to discuss architecture details, security requirements, or deployment strategy

Powered by Claude Agent SDK
Use ← → arrow keys or click to navigate