Built for Atlas and Titan Intelligence — Agents SDK + Durable Objects + Workflows, the stack behind the agentic OS for the trades
QBR FOLLOW-UP — AGENTIC PLATFORM

The agentic OS for the trades, on Cloudflare.

On the Q4 earnings call, Vahe said FY27 is about "bringing the Agentic Operating System for the trades to life" and doubling the capacity of Max. That platform needs three primitives: stateful agents, durable orchestration, and edge-grade inference. We ship all three.

Agents SDK Durable Objects Workflows Workers AI AI Gateway Realtime
Live sandbox uses real Workers, R2, D1, Vectorize, and Workers AI
atlas request · agent path
// user asks Atlas: "book the heat-pump install"
Workerrequest enters edge · 8ms
→ Durable Objectper-user agent state
→ Agents SDKtool selection · streaming
→ AI Gatewaycache · fallback · audit
→ Workers AI / OpenAIreasoning step
→ Vectorizeper-tenant pricebook RAG · <100ms
→ Workflowbook · charge · dispatch · text
→ D1 + Hyperdriveoperational state · Postgres bridge
Statefulper user, per tenant
Durablecrash-safe orchestration
Agents SDK + Durable Objects — the platform behind Atlas

ServiceTitan · NASDAQ: TTAN · FY26 results (year ended Jan 31, 2026)

10,800+
Active Customers
$82B
Gross Transaction Volume
$1B+
Annualized Run Rate
>110%
Net Dollar Retention
LIVE — NO SIGNUP REQUIRED

Skip the slide deck.
Try it live.

The search above is a preview. The Live Sandbox is the per-tenant inference + retrieval layer that sits behind Atlas — upload sample tenant content (pricebook page, service-agreement PDF, technician photo), watch Workers AI extract structure, see Vectorize index it, then query with natural language. Same primitive, every tenant, isolated by Workers for Platforms.

Your own isolated session
Real Workers AI inference
Reset anytime
sandbox-api.workers.dev
POST /api/upload
→ R2 20ms
→ Workers AI caption 1.2s
→ BGE embed 60ms
→ Vectorize index 50ms
POST /api/search
→ Embed query 60ms
→ Vectorize query 12ms
→ D1 lookup 5ms
Total ~77ms

A ServiceTitan tenant's world, on the edge

Each artifact lives in a per-tenant R2 bucket, gets embedded by Workers AI, and is queryable from any Atlas conversation through its Durable Object — sub-100ms, regardless of where the user is.

Service Agreement
HVAC · Annual
Customer Doc
Membership · Gold Tier
Renews Q3 · $389/yr
Dispatch Job
Emergency · Heat Out
Live Operations
3-Hour Window · 4pm
ETA optimized via Dispatch Pro
Pricebook Entry
Heat Pump · 3-Ton
Catalog
Trane XR16 Install
$8,450 · 1-day install
Technician
Master · HVAC
Workforce
Mike R. · 12 yrs
4.9★ · 92% first-call fix
HERO LAYER — STATEFUL AGENTS

Per-tenant agent state, two ways

Atlas is a tool-using agent that lives per user, per tenant. The hard problem isn't the LLM call — it's the state, the orchestration, and the retrieval around it. Cloudflare gives you two composable primitives.

Live conversation → Durable Objects

One DO per active Atlas conversation (or per Voice Agent call). Strongly consistent, single-threaded, edge-resident. Holds the agent's working memory, tool-call history, and session tokens — without Redis, sticky load balancing, or a separate session service.

Durable Objects + Agents SDK · purpose-built for stateful agents

Grounded knowledge → AI Search + Vectorize

Drop tenant documents (pricebooks, service agreements, technician manuals, regional regs) into R2. AI Search auto-generates and maintains a per-tenant Vectorize index. No embedding pipeline to build, no ETL to maintain.

R2 + AI Search + Vectorize · per-tenant out of the box
Your existing stack stays put. Postgres / MySQL of record stay where they are — Hyperdrive bridges from Workers. AI Gateway sits in front of your existing OpenAI / Anthropic / Google spend, so nothing about your provider relationships has to change to start.

Architecture: ServiceTitan on Cloudflare

Every Titan Intelligence workload mapped to a Cloudflare primitive. The three ringed in orange are the agentic platform foundation.

Agent Framework — Agents SDK

Opinionated framework for Atlas: tool-use, streaming, scheduling, hooks, memory. Your team builds trade-specific tools; we own the orchestration plumbing.

THE PLATFORM PRIMITIVE

Agent State — Durable Objects

One DO per user / tenant / conversation. Strongly consistent, sticky, edge-resident. The memory and concurrency model Atlas and the Voice Agents need.

STATEFUL EDGE

Orchestration — Workflows

Durable, retryable, long-running. A Voice Agent call that books a job, charges a card, dispatches a tech, and texts the homeowner shouldn't break on a partial failure.

DURABLE STEPS

Edge Inference — Workers AI

Sidecar work on the voice path: intent routing, PII redaction, transcript summarization. Llama / Mistral / Whisper / BGE — pay per use, no GPUs to run.

Voice latency · edge sidecar

LLM Control Plane — AI Gateway

Cost per Titan Intelligence feature, prompt caching for high-overlap Atlas system prompts, provider fallback when OpenAI has an outage, full audit. Margin lever for FY27.

Across your existing OpenAI / Anthropic / Google

Voice — Realtime / WebRTC

For Karen, Lucy, Piper, Jenny. Low-latency audio + server-side processing at the edge. Pairs with Workers AI for the sub-800ms voice-turn budget.

Voice Agent Pro · Contact Center Pro

Tenant Knowledge — Vectorize + AI Search

Per-tenant vector indexes over pricebooks, service agreements, technician manuals. AI Search handles ingest → chunk → embed → retrieve so engineering ships features, not pipelines.

Per-tenant out of the box

Operational Store — D1 + Hyperdrive

D1 for agent-local operational data (transcripts, tool-call traces, session config). Hyperdrive pools connections to your existing Postgres / MySQL system of record.

Your DB of record stays put

Already-Paid Capability — Bot Mgmt + WAF + API Shield

Per the 4/22 TAR: proxy api.servicetitan.com, move WAF rules out of Log mode on top paths, rate-limit auth. API Shield is the natural next step once endpoints are proxied.

Turn on what you already own

The business case

FY27 guidance: revenue growth 24% → ~16%, operating margin expanding meaningfully. Efficiency matters. AI cost goes from a line item to a P&L driver. These three primitives put Cloudflare directly on the levers ServiceTitan is being measured against publicly.

<800ms
Voice-Turn Budget

Workers AI sidecar work + Realtime WebRTC at the edge keeps Karen / Lucy / Piper / Jenny inside the natural-conversation latency window as you scale tenants.

1
AI Control Plane

AI Gateway in front of every Titan Intelligence call. Cost per feature, cost per tenant, prompt caching, provider fallback. Direct margin lever for the FY27 efficiency narrative.

$82B
GTV Atlas Can Reason Over

Stateful, per-user, per-tenant agent infra at the edge. Atlas grounded in pricebook + service-agreement + dispatch context — at $1B+ run-rate scale.

Mapped to the strategic commitments you've already made publicly

"Agentic Operating System for the trades." Agents SDK + Durable Objects + Workflows + Workers. The platform behind the platform.
"Doubling the capacity of Max." Workers AI for edge inference, AI Gateway for cost/observability, Realtime for voice. Capacity that scales horizontally without GPU operations.
"$1B run rate, expanding margins." AI Gateway prompt caching + provider fallback are direct margin levers on multi-provider LLM spend.
"Democratize automation in the trades." Edge compute makes Atlas and the Voice Agents fast and cheap enough to ship to the smallest contractor on a phone in a truck.
"Step-function change in velocity." Workers deploy speed + Agents SDK = days-not-quarters product velocity for new trade-specific agents.
Already-paid security (4/22 TAR). Proxy api.servicetitan.com, exit Log-only WAF mode on top paths, rate-limit auth. Value extraction before any new contract.

Where to next?

A 60-minute working session: your Atlas and Voice Agent technical leads with our Agents SDK and AI Gateway PMs. No slides, whiteboard. Concrete numbers — not architecture diagrams.

Built with Cloudflare Pages, Agents SDK, Durable Objects, Workflows, Workers AI, AI Gateway, Realtime, Vectorize, AI Search, D1, Hyperdrive, R2, and the security stack you already pay for.
Deployed globally to 330+ cities. Your DB of record and provider relationships unchanged.