A presentation at Conf42 AI Agents 2026 by Kemet Dugue
Integrating Agentic Workflows Scaling Productivity for Uber’s Driver Onboarding Team Kemet Dugue • Fullstack Software Engineer @ Uber kemetdugue.com • linkedin.com/in/kemet-dugue CONF42 AI AGENTS 2026
OVERVIEW Agenda 1. Introduction & Domain Landscape 2. Developer Workspace Skills 3. Tooling & Platform Architecture 4. Major Case Study & Demo: Driver Test Account AI Agent 5. Impact & Key Takeaways CONF42 AI AGENTS 2026
INTRODUCTION & DOMAIN LANDSCAPE Driver Onboarding Compliance & Screenings Earner Lifecycle Management Multi-Channel Onboarding Identity & Document Verification CONF42 AI AGENTS 2026
CASE STUDY DEVELOPER WORKSPACE Friction: Enabling True Agentic Development with Monorepo Scale Fix: Isolated /worktree Automation Targeted Monorepo Sparse-Checkout The custom /worktree command isolates the Driver Onboarding sub-tree instantly, eliminating indexing overhead and branch collisions. Zero full-repo re-indexing penalties 2,200 Services Parallel multi-task development across clean workspaces Massive unified monorepo causes checkout lag, indexing slowdowns, and frequent local config drift during branch switching.
CONF42 AI AGENTS 2026
CASE STUDY DEVELOPER WORKSPACE Friction: Strict CI Gates Fix: Branch-to-Diff Skill Branch-to-Diff Rule Strict CI Gatekeeping Extracts Jira context, verifies local unit test coverage, and formats PR Rigid PR templates and mandatory test summaries matching team guidelines to pass CI on attempt #1. evidence blocks cause verbose, generic LLM diffs to fail formatting gates repeatedly. CONF42 AI AGENTS 2026 Pass CI validation on first attempt
CASE STUDY DEVELOPER WORKSPACE Friction: Knowledge Drift Fix: EngWiki Sync Skill EngWiki Doc Sync Knowledge Drift Synchronizes Confluence and internal guides directly with updated Outdated documentation leads code models service schemas, eliminating outdated patterns in LLM completions. to generate obsolete testing signatures and deprecated mock configurations. CONF42 AI AGENTS 2026 Ground truth documentation for agents
CASE STUDY DEVELOPER WORKSPACE Friction: Administrative Toil 📋 Fix: JIRA REST Skill JIRA REST Skill Manual JIRA Tracking Automated ticket creation with team namespaces, pulls acceptance Engineers spend valuable sprint time criteria into prompts, and transitions ticket states manually creating subtasks, filling required namespace fields, and synchronizing ticket status after diff merges CONF42 AI AGENTS 2026
TOOLING & PLATFORM ARCHITECTURE Uber’s AI Platform & Ecosystem Agent Builder MCP Gateway Source: Pragmatic Engineer CONF42 AI AGENTS 2026
TOOLING & PLATFORM ARCHITECTURE Agent Builder ● No Code Agent ● Native Tooling & MCP Integration ● Dynamic Sub-Agent Handoffs Source: Pragmatic Engineer CONF42 AI AGENTS 2026
TOOLING & PLATFORM ARCHITECTURE MCP Gateway ● Exposes Core Microservices as Tools ● Decouples Agent Logic from Service APIs ● Built in Observability & Audit Trails Source: Pragmatic Engineer CONF42 AI AGENTS 2026
MAJOR CASE STUDY TEST ACCOUNT AGENT Test Account AI Agent 15,000+ cities 15 Background Check Error States San Francisco, Seattle, Accra, Insufficient Licensing History, London, NYC, Bangalore… failed SSN, General Rejection… 20+ earning opportunities 10 Document Rejection States Car Seat, Alcohol Delivery, Poor Picture Quality, Wrong Eats, Rides, Tasks, Commercial Document, Document Blurry… Rides… Vehicle Insurance States Terms and Conditions states CONF42 AI AGENTS 2026 Various Surfaces New features.. → 50,000 Unique Onboarding States Uploaded Driver License, Rejected Profile Photo, Expired Document, Incomplete T&Cs
MAJOR CASE STUDY TEST ACCOUNT AGENT Test Account AI Agent - Internals Create a delivery test account in Seattle with an approved driverʼs license but failed background check Phase 1 Create Test Account Phase 2 Transition Test Account (to proper state) CONF42 AI AGENTS 2026
IMPACT & KEY TAKEAWAYS Takeaways Quantifiable Business & Engineering Impact 30s 01 Decouple Reasoning from State Mutation Use agents for dynamic discovery, but hand off execution to atomic batch engines. Setup Duration Down from 510 minutes 02 Let Agents Work in Parallel Use git worktrees or other mechanisms so multiple agents can work at the same time without blocking each other. 95% Creation Success Rate Across benchmark test suite 100% Self-Serve Access Via Slack BMO Bot for QA & Ops CONF42 AI AGENTS 2026 03 Benchmark Tool Calls, Not Vibe Checks 04 Safety Belongs in Code, Not Prompts Skip subjective LLM judges; test agent reliability using deterministic tool selection accuracy and task coverage. Never rely on prompt instructions for security; block invalid mutations with deterministic backend interceptors.
THANK YOU! github.com/kdugue CONF42 AI AGENTS 2026 kemetdugue.com linkedin.com/in/kemet-dugue
In this session, we’ll dive into a case study on how the Driver Onboarding team transitioned lots of manual, script-heavy engineering processes into agentic-based ones. We’ll explore how various AI agents were integrated into the developer lifecycle to scale team output. The session will include: - Existing workflows and the pain points associated with them - How the team transitioned those workflows into agentic-based ones - Overview of the AI-based tools and landscape at Uber - Best practices for integrated agentic workflows
for free. You
can too.