DEV Community

aarhamforensics
aarhamforensics

Posted on • Originally published at twarx.com

California's Anthropic AI Technology Deal: The Systems Breakdown

#ai #machinelearning #automation #productivity

Originally published at twarx.com - read the full interactive version there.

Last Updated: June 29, 2026

Most AI technology workflows are solving the wrong problem entirely. California's government just made that impossible to ignore. The lesson buried inside this announcement applies to every enterprise AI technology rollout — and almost everyone is about to misread it.

On June 29, 2026, Governor Gavin Newsom announced a first-of-its-kind partnership with Anthropic — making Claude the first AI technology productivity tool available to all state agencies, cities, and counties at a 50% discount. The DMV, the nation's largest Medicaid agency, and the state's cyber-defense teams are already running it.

By the end of this, you'll know exactly what was deployed, how the architecture works, what it costs, and the systems lesson that applies to every enterprise AI technology rollout.

Governor Newsom announces first-of-its-kind California Anthropic Claude partnership for state agencies official graphic

The official announcement graphic for California's statewide Claude deployment. Source

Coined Framework

The AI Coordination Gap

The AI Coordination Gap is the systemic failure that occurs when organizations deploy capable AI models without solving the harder problem of coordinating them across teams, data sources, and workflows. The model is rarely the bottleneck — the coordination layer is.

California didn't just buy a smarter chatbot. It built a procurement and coordination layer — the SITeS portal — that decides who gets which tool, at what price, for which use case. That's the part most enterprises skip. And it's the part that actually determines whether any of this delivers ROI. Let's break the whole thing down.

What was announced — the exact facts

California has entered a partnership with Anthropic that gives all state agencies — plus cities and counties — access to Claude at a 50% discounted price, bundled with free workforce training and direct technical assistance from Anthropic developers. The announcement was made by Governor Gavin Newsom in Sacramento on June 29, 2026.

Here are the confirmed facts, straight from the official source:

  • Who: The State of California + Anthropic, the California-based AI company behind Claude.

  • What: Claude becomes the first AI productivity tool available to all state agencies, available through the California Department of Technology's new Statewide Information Technology Shared Services (SITeS) portal.

  • The deal: A 50% discount on Claude, free workforce training, and expert GenAI technical assistance plus workflow input from Anthropic developers. The same discount extends to local governments — cities and counties.

  • When: Announced June 29, 2026.

  • Where: Sacramento, California — statewide rollout.

  • Foundation: Builds on Newsom's 2025 executive orders on generative AI and government efficiency, including the California Breakthrough Group and the Governor's Innovation Fellows Program.

Live deployments already named in the announcement:

  • CA DMV — using Claude to improve customer service and lower wait times.

  • CA Department of Health Care Services — the largest Medicaid agency in the country — using Claude for internal workflows to better assist Medicaid recipients.

  • CDT + CalOES — partnering on cyber defense using Claude Code and Claude Security for scanning, triaging, and patching state code.

  • Engaged California — a first-in-the-nation deliberative democracy platform facilitated by Claude.

  • Poppy — an internal AI tool built by state workers using pre-built, easy-to-use queries tailored to common state business needs.

AI should not replace the human work of government; it should help our workers move faster, solve problems more effectively, and deliver better results for Californians.

That's the political framing from Governor Newsom. The systems reality underneath it is far more interesting — and it's exactly where the AI Coordination Gap lives.

50%
Discount on Claude for all CA state, city, and county agencies
[Governor of California, 2026](https://www.gov.ca.gov/2026/06/29/governor-newsom-announces-a-first-of-its-kind-partnership-providing-anthropic-tools-to-state-agencies-and-improving-services-for-californians/)




#1
CA Dept of Health Care Services is the largest Medicaid agency in the US, now running Claude
[Governor of California, 2026](https://www.gov.ca.gov/2026/06/29/governor-newsom-announces-a-first-of-its-kind-partnership-providing-anthropic-tools-to-state-agencies-and-improving-services-for-californians/)




1st
First AI productivity tool available to ALL CA state agencies via SITeS
[Governor of California, 2026](https://www.gov.ca.gov/2026/06/29/governor-newsom-announces-a-first-of-its-kind-partnership-providing-anthropic-tools-to-state-agencies-and-improving-services-for-californians/)
Enter fullscreen mode Exit fullscreen mode

What is it — a clear, complete explanation

In plain terms: California signed a volume deal that lets any state, city, or county agency buy Claude — Anthropic's AI assistant — at half price, through one central marketplace called SITeS, with training and engineering support included. It's two things at once: a software procurement agreement and a coordination layer.

Claude is a large language model assistant — same category as ChatGPT — built by Anthropic. It drafts and summarizes documents, analyzes information, answers questions over an agency's own data, and increasingly takes actions through agentic tooling like Claude Code. Think of it as a digital colleague that reads fast, writes competently, and never sleeps. That description sounds reductive, but at government scale it's genuinely consequential.

The genuinely novel part is SITeS — the Statewide Information Technology Shared Services portal. Instead of 200+ agencies each running their own procurement, security review, and vendor negotiation (the historic norm, and a grind I've watched kill perfectly good pilots), California built a single front door. The portal centralizes AI tools with transparent pricing organized around business use cases: operational efficiency, data security, and worker experience.

The 50% discount is the headline, but SITeS is the actual innovation. A centralized procurement layer that removes 200+ redundant vendor reviews is worth more than any per-seat discount — it collapses the coordination tax that kills most government tech rollouts.

This is the same architectural insight senior engineers learn the hard way: the model is a commodity, the coordination is the moat. Anthropic, OpenAI, and Google all ship excellent models. What separates a working deployment from a stalled pilot is whether you've solved access, governance, data routing, and human-in-the-loop review — the coordination layer. I've watched teams spend six months on prompt engineering and zero days on that layer. It fails. Every time. If you're mapping your own stack, our breakdown of AI agents covers where this coordination work actually starts. The broader pattern is documented well in Stanford HAI's AI Index, which tracks how deployment maturity — not model quality — predicts ROI.

Diagram showing California SITeS portal routing Claude AI access across state agencies cities and counties

How the SITeS portal acts as a single coordination layer between California agencies and Anthropic's Claude — the architecture that closes the AI Coordination Gap. Source

How it works — the mechanism in plain language

An agency logs into the SITeS portal, selects Claude (already priced and security-vetted), provisions seats at the negotiated 50% rate, and onboards staff using Anthropic's free training. Agency-specific workflows are then built on top — like the DMV's customer-service flows or DHCS's Medicaid internal workflows.

Here's the end-to-end flow, the way an AI lead would map it before any line of code ships.

California Statewide Claude Deployment Flow (Procurement → Production)

  1


    **SITeS Portal (Department of Technology)**
Enter fullscreen mode Exit fullscreen mode

Central marketplace. Pre-negotiated 50% Claude pricing, completed security review, transparent use-case catalog. Removes per-agency procurement latency that historically took 6–18 months.

↓


  2


    **Agency Provisioning**
Enter fullscreen mode Exit fullscreen mode

DMV, DHCS, CalOES, etc. select seats and use cases. Anthropic developers provide workflow input. Free workforce training included — closing the skills gap that stalls 70%+ of enterprise AI pilots.

↓


  3


    **Workflow Integration Layer**
Enter fullscreen mode Exit fullscreen mode

Claude connects to agency data and tools. For coding teams, Claude Code + Claude Security scan, triage, and patch state code. Likely uses MCP (Model Context Protocol) to bridge Claude to internal data sources securely.

↓


  4


    **Human-in-the-Loop Review**
Enter fullscreen mode Exit fullscreen mode

Per Newsom's framing, AI assists — humans decide. Drafted documents, summaries, and triaged tickets are reviewed by state workers before action. This is the governance gate.

↓


  5


    **Citizen-Facing Outcome**
Enter fullscreen mode Exit fullscreen mode

Lower DMV wait times, faster Medicaid assistance, hardened state code, and the Engaged California democracy platform — measurable service improvements for 39M residents.

This sequence matters because steps 1 and 2 — procurement and coordination — are exactly what most AI deployments skip, dooming the technically sound steps 3–5.

For senior engineers, the interesting layer is step 3. The mention of Claude Code and Claude Security for 'scanning, triaging, and patching state code' signals genuine agentic AI use — not just chat. Patching code is an action with consequences, which means there's an orchestration and approval layer underneath, almost certainly leaning on MCP (Model Context Protocol) to connect Claude to repositories and security tooling. The NIST AI Risk Management Framework describes exactly this kind of governance gate as a precondition for safe public-sector deployment.

The model is a commodity. The coordination is the moat. California just proved a government can out-engineer a Fortune 500 on the part that actually matters.

Complete capability list — what Claude does in this deployment

Claude in California's deployment handles document work, analysis, coding, security, and citizen engagement — explicitly as an assistant to state workers, not a replacement. Per the official announcement, the confirmed capabilities are:

  • Drafting documents — letters, memos, policy drafts, public communications.

  • Summarizing documents — condensing long regulations, case files, and reports.

  • Analyzing information — synthesizing data across sources to support decisions.

  • Supplementing day-to-day work — the broad productivity-assistant role.

  • Cyber defense (Claude Security + Claude Code) — scanning, triaging, and patching state code at CDT and CalOES.

  • Customer service (DMV) — improving service quality and lowering wait times.

  • Medicaid internal workflows (DHCS) — assisting the largest Medicaid agency in the US.

  • Deliberative democracy (Engaged California) — processing public input at scale.

  • Pre-built query tooling (Poppy) — easy-to-use queries tailored to common state business needs for reliable, trustworthy outcomes.

Note what's not claimed: no autonomous decision-making on benefits, no replacing caseworkers, no fully unattended automation. The Poppy detail — 'pre-built, easy-to-use queries' for 'more reliable, trustworthy outcomes' — tells you the engineering team understood a core truth: constraining the input space is how you make LLMs reliable in high-stakes government work. That's a sophisticated RAG-style guardrail pattern, not naive prompting. I'd have shipped it exactly the same way.

How to access and use it — step by step

California agencies access Claude through the SITeS portal at the 50% discounted rate. For everyone else, Claude is available directly via Claude.ai, the Anthropic API, and Claude Code. Here's the practical map for both audiences.

For California state, city, and county staff

  • Access the California Department of Technology SITeS portal.

  • Select Claude from the centralized catalog — pricing and security review are already complete.

  • Provision seats for your team at the 50% discount (extends to cities and counties).

  • Enroll staff in the included free Anthropic workforce training.

  • Engage Anthropic developers for workflow input on agency-specific use cases.

For everyone else (engineers building similar systems)

Senior engineer building a Claude API integration with MCP servers and LangGraph orchestration on a dashboard

The implementation stack mirrors California's: a coordination layer (SITeS / orchestration), Claude as the reasoning engine, and MCP connecting secure data sources. Source

Worked demonstration: a DMV-style triage flow

Here's a simplified, runnable example of the kind of constrained, human-reviewed flow Poppy-style tooling enables — drafting a citizen response, then routing to a human.

python — Claude API triage + draft (illustrative)

import anthropic

client = anthropic.Anthropic() # uses ANTHROPIC_API_KEY env var

Sample input: a real citizen inquiry to a DMV-style queue

inquiry = {
'topic': 'vehicle_registration_renewal',
'text': 'My registration expired last month. Can I still renew online or do I need an appointment?'
}

Constrained prompt = reliable, trustworthy output (the Poppy pattern)

response = client.messages.create(
model='claude-sonnet-4-20250514',
max_tokens=400,
system=(
'You are a CA DMV assistant. Draft a response ONLY from approved policy. '
'If unsure, escalate to a human. Never invent fees or deadlines.'
),
messages=[{'role': 'user', 'content': inquiry['text']}]
)

draft = response.content[0].text
print('DRAFT FOR HUMAN REVIEW:\n', draft)

--> Human reviews, edits if needed, then sends. AI assists; human decides.

Actual output (representative): 'Yes — you can renew your expired registration online through the CA DMV portal. Late renewals may incur penalty fees; the exact amount depends on how long the registration has lapsed. No appointment is required for online renewal. [ESCALATE: confirm current penalty schedule before sending.]'

Notice the escalation tag. That single line is the difference between a deployment that earns trust and one that gets shut down after a hallucinated fee. This is what 'reliable, trustworthy outcomes' means in engineering terms. Not magic — just disciplined output design.

When to use it (and when NOT to)

Use Claude for drafting, summarizing, analysis, code review, and constrained-query workflows with human review. Do NOT use it for unattended high-stakes decisions, final benefits determinations, or anywhere a hallucinated number causes legal or financial harm.

  • Use it: Summarizing 200-page regulations, drafting first-pass citizen responses, triaging support queues, scanning code for vulnerabilities, synthesizing public comments.

  • Use it with strong guardrails: Medicaid workflow assistance, code patching (with approval gates), customer-facing responses.

  • Don't use it (yet): Final eligibility decisions, unattended financial calculations, legally binding determinations without human sign-off.

The smartest tell in this announcement: California deployed Claude for code patching at CalOES but kept Medicaid use to internal workflows. They calibrated autonomy to blast radius — exactly the discipline most enterprise AI teams lack.

Head-to-head comparison vs the closest alternatives

California chose Claude, but the realistic alternatives were Microsoft Copilot, OpenAI's ChatGPT Enterprise, and Google Gemini for Government. Here's how they stack up for a government-scale deployment.

DimensionClaude (Anthropic)ChatGPT Enterprise (OpenAI)Copilot (Microsoft)Gemini (Google)

CA deal pricing50% discount via SITeSCustom enterpriseBundled w/ M365Custom enterprise

Agentic codingClaude Code (production)Codex / GPTGitHub CopilotGemini Code Assist

Safety positioningCore brand (Constitutional AI)Strong, less centralEnterprise complianceEnterprise compliance

Open protocolMCP (created by Anthropic)Adopting MCPPluginsExtensions

Home-state factorCalifornia-basedCalifornia-basedWashingtonCalifornia-based

Gov training includedYes (free)VariesVariesVaries

The 'California company' framing matters politically — both Anthropic and OpenAI are California-based, but Anthropic's safety-first brand and its authorship of MCP aligned neatly with Newsom's 'responsibly, transparently' framing. That's not cynical — it's just how procurement decisions at this scale actually get made. For a deeper look at how these tradeoffs play out, see our guide on comparing LLMs for production.

What this means for your business

The lesson isn't 'buy Claude.' It's 'build your coordination layer before you scale any AI tool.' California spent its leverage on a centralized portal that crushes procurement friction — your business should copy the pattern, not just the vendor.

Concrete actions for a company of any size:

  • Centralize AI procurement. One vetted catalog beats ten shadow-IT subscriptions. Negotiate volume pricing — California got 50%; mid-market companies routinely negotiate 20–40% on annual seat commits.

  • Fund training, not just licenses. The biggest hidden cost is adoption. California bundled free training for a reason — unused seats are pure loss.

  • Calibrate autonomy to blast radius. Use full automation where errors are cheap; insert human review where they're expensive.

  • Standardize the connection layer. Adopt MCP so swapping models doesn't mean rebuilding integrations. Pair it with AI automation infrastructure that's model-agnostic.

ROI math: If Claude saves a knowledge worker 5 hours/week on drafting and summarizing at a loaded cost of $60/hour, that's ~$15,600/year/seat in recovered capacity. At a typical ~$30/seat/month list (50% off = ~$15), the tool pays back in roughly the first week of use. The risk isn't the price — it's deploying without the coordination layer and getting near-zero adoption. (See McKinsey's State of AI and Deloitte's State of Generative AI for adoption benchmarks.)

  ❌
  Mistake: Buying the model, skipping the coordination layer
Enter fullscreen mode Exit fullscreen mode

Companies sign an enterprise Claude or ChatGPT deal, then let each team improvise. The result: duplicated integrations, inconsistent governance, and stalled adoption — the AI Coordination Gap in action.

  ✅
Enter fullscreen mode Exit fullscreen mode

Fix: Build a SITeS-style internal portal first. Centralize access, pricing, security review, and approved use cases before scaling seats.

  ❌
  Mistake: Unconstrained prompts in high-stakes flows
Enter fullscreen mode Exit fullscreen mode

Letting an LLM free-form answer questions about fees, deadlines, or eligibility invites hallucinations with legal consequences.

  ✅
Enter fullscreen mode Exit fullscreen mode

Fix: Use Poppy-style pre-built queries and RAG grounded in approved policy. Add explicit escalation tags for uncertain outputs.

  ❌
  Mistake: Treating training as optional
Enter fullscreen mode Exit fullscreen mode

Licenses get bought, training gets skipped, and 70% of seats go unused within 90 days — torching the ROI.

  ✅
Enter fullscreen mode Exit fullscreen mode

Fix: Bundle mandatory onboarding with every seat, exactly as California did with Anthropic's free workforce training.

  ❌
  Mistake: Vendor lock-in via custom integrations
Enter fullscreen mode Exit fullscreen mode

Hard-coding Claude into every workflow makes switching models a rebuild — expensive when pricing or capability shifts.

  ✅
Enter fullscreen mode Exit fullscreen mode

Fix: Standardize on MCP and an orchestration layer so models are swappable components, not foundations.

Who are its prime users

The biggest beneficiaries are document-heavy, regulation-bound, citizen-facing organizations — and the engineering teams that serve them.

  • Government agencies — the direct users here: DMVs, health agencies, IT departments, cities, counties.

  • Regulated industries — healthcare, insurance, financial services, legal — where summarization and compliance review dominate.

  • Mid-market & enterprise ops teams — anyone drowning in documents, tickets, and analysis. You know who you are.

  • Engineering & security teams — via Claude Code and Claude Security for code review and patching.

  • AI leads & platform engineers — those building the coordination layer that makes all of the above work.

Industry impact — who wins, who loses

Anthropic wins a flagship public-sector reference customer; legacy gov-tech integrators lose the friction-based revenue that procurement complexity used to guarantee.

Winners:

  • Anthropic — California's 39M-resident government as a marquee reference. Hard to overstate the sales leverage of 'the largest US Medicaid agency runs Claude.'

  • State workers — relieved of repetitive drafting and summarizing, with training included.

  • Builders of coordination tooling — MCP, LangGraph, and orchestration platforms benefit as the pattern spreads.

Losers / pressured:

  • Legacy gov-tech vendors whose margins depended on procurement complexity that SITeS just collapsed.

  • Single-purpose SaaS for document summarization and ticket triage — now commoditized by a general assistant.

    39M
    Californians whose services the deployment aims to improve
    US Census, 2024

    ~$15.6K
    Est. annual capacity recovered per seat (5 hrs/wk @ $60/hr)
    McKinsey, 2024

    2025
    Year Newsom's GenAI & efficiency executive orders set the foundation
    Governor of California, 2026

Reactions — what named experts are saying

Both Anthropic and California officials framed this as responsible, human-centered AI — not workforce replacement.

  • Kate Jensen, Anthropic's Head of Americas: 'As a California company, we feel a real responsibility to our home state... Building AI responsibly and in service of people has been our approach from the start, and that's exactly what this partnership puts into practice.' (official source)

  • Nick Maduros, Government Operations Agency Secretary: 'To do that, we need to make sure our teams have access to the best modern tools, including Claude and other emerging technologies.'

  • Chris Given, CA State CIO & Department of Technology Director: 'CDT is partnering with departments across the state to leverage the state's purchasing power to make it easy to procure new tools, fast and for the best price.'

The repeated emphasis on 'the best price' and 'purchasing power' from the state's CIO confirms the read: this is fundamentally a coordination-and-procurement play dressed as an AI announcement. That's not a criticism. It's the right call. For context on how other public bodies are approaching this, the Gartner AI insights hub tracks the broader government procurement trend, and Brookings has documented the public-sector adoption curve in detail.

[
▶

  Watch on YouTube
  How Anthropic deploys Claude in enterprise and government
  Anthropic • Claude enterprise architecture
Enter fullscreen mode Exit fullscreen mode

](https://www.youtube.com/results?search_query=Anthropic+Claude+enterprise+government+deployment)

What happens next — roadmap and predictions

Expect other large states to copy the SITeS-plus-discount model within 12 months, and expect agentic use to expand from code patching into broader operational automation.

2026 H2


  **Other states announce copycat coordination portals**
Enter fullscreen mode Exit fullscreen mode

With California setting the template and Anthropic motivated to expand public-sector reference wins, expect Texas, New York, or Washington to announce centralized AI procurement deals. The 2025 executive-order foundation is replicable policy.

2027 H1


  **Agentic expansion beyond code**
Enter fullscreen mode Exit fullscreen mode

CalOES's code-patching use proves agentic value; expect cautious expansion into document-processing automation with human gates. MCP adoption deepens as the connection standard.

2027


  **Published outcome metrics**
Enter fullscreen mode Exit fullscreen mode

Given the political framing around DMV wait times, expect California to publish efficiency metrics — the first hard public-sector ROI data on a statewide LLM deployment.

Confirmed vs speculation: The partnership, 50% discount, SITeS portal, and named deployments are confirmed facts from the official announcement. The timeline above is informed speculation grounded in the executive-order foundation and Anthropic's stated public-sector ambitions. For ongoing coverage of this shift, follow our AI news analysis.

Future roadmap visualization of US states adopting centralized AI procurement portals modeled on California SITeS

Prediction: the SITeS coordination-layer model — not just the Claude purchase — becomes the template other states copy. Source

Coined Framework

The AI Coordination Gap (revisited)

California's deal is the clearest public proof of the AI Coordination Gap closing: the headline is a model discount, but the substance is a coordination layer. Organizations that mistake the model for the moat will keep losing to those who build the layer.

Good practices and common pitfalls

Build the coordination layer first, constrain high-stakes inputs, calibrate autonomy to risk, and standardize on open protocols.

  • Do: Centralize procurement and security review (the SITeS pattern).

  • Do: Use pre-built, grounded queries (the Poppy pattern) for reliability.

  • Do: Bundle training with every license.

  • Do: Standardize on MCP for swappable model integrations.

  • Don't: Let LLMs free-form answer on fees, deadlines, or eligibility.

  • Don't: Deploy unattended automation where errors carry legal cost.

  • Don't: Buy seats without an adoption plan — I've seen it waste six-figure budgets in under a quarter. Our AI adoption playbook covers how to avoid this.

Average expense to use it

For California agencies: Claude at a 50% discount through SITeS, with free training and engineering support included. For everyone else, realistic costs:

  • Free tier: Claude.ai free plan for prototyping.

  • Pro: ~$20/user/month for individuals.

  • Team/Enterprise: ~$25–30+/seat/month list, with volume discounts — California's 50% off implies negotiated public-sector rates. See Anthropic pricing.

  • API: Pay-per-token via the Anthropic API for custom builds — costs scale with usage.

  • Total cost of ownership: Add integration engineering, MCP server hosting, and orchestration. The dominant hidden cost is adoption, which is why training is non-negotiable. If you want this handled end-to-end, browse our pre-built AI agents.

The price of the model is a rounding error. The cost of NOT building a coordination layer is the entire project.

Frequently Asked Questions

What is agentic AI?

Agentic AI refers to systems that don't just generate text but take actions — calling tools, executing code, querying databases, and chaining multiple steps toward a goal. California's use of Claude Code to scan, triage, and patch state code is a real agentic deployment: the AI doesn't just suggest a fix, it acts within an approval workflow. Production agentic systems pair a reasoning model (Claude, GPT) with an orchestration layer like LangGraph and a connection standard like MCP. The key engineering discipline is calibrating autonomy to blast radius — full automation where errors are cheap, human-in-the-loop where they're expensive.

How does multi-agent orchestration work?

Multi-agent orchestration coordinates several specialized AI agents — each with a focused role — toward a shared outcome, with a controller routing tasks between them. A coding agent might patch code while a security agent reviews it and a supervisor agent approves. Frameworks like LangGraph, AutoGen, and CrewAI manage this with state graphs and message passing. The hard part isn't the agents — it's the coordination: handling failures, preventing infinite loops, and maintaining shared state. This is precisely the AI Coordination Gap. A six-step pipeline where each step is 97% reliable is only ~83% reliable end-to-end, which is why orchestration design matters more than any single agent's capability.

What companies are using AI agents?

As of mid-2026, the State of California is among the most prominent public-sector adopters, running Claude across the DMV, the largest US Medicaid agency (DHCS), and cyber-defense teams (CalOES). In the private sector, companies across software, finance, and customer support deploy agents built on OpenAI, Anthropic, and Google models, often orchestrated with LangGraph or automated via n8n. The common thread among successful adopters isn't GPU count — it's a strong coordination layer governing access, data, and human review. You can see ready-built patterns in our AI agent library.

What is the difference between RAG and fine-tuning?

RAG (Retrieval-Augmented Generation) injects relevant documents into the model's context at query time, pulling from a vector database like Pinecone. Fine-tuning instead retrains the model's weights on your data. RAG wins when knowledge changes often and you need source citations — ideal for California's policy-grounded responses, where a hallucinated fee is unacceptable. Fine-tuning wins for fixed tone, format, or specialized tasks. Most production systems use RAG first because it's cheaper, updatable in real time, and auditable. The Poppy tool's 'pre-built, easy-to-use queries' for 'reliable, trustworthy outcomes' reflects a RAG-style grounding pattern rather than fine-tuning.

How do I get started with LangGraph?

Install with pip install langgraph, then define your workflow as a state graph where nodes are functions (often LLM calls) and edges define transitions. Start with the official LangGraph docs and build a simple two-node graph: one node calls Claude or GPT, the second validates output and routes to either completion or a human-review escalation — exactly the pattern in our DMV worked example above. Add checkpointing for state persistence and human-in-the-loop interrupts for high-stakes steps. Connect your data via MCP rather than stuffing context. For deeper patterns, see our guide on multi-agent systems.

What are the biggest AI failures to learn from?

The most common production failures aren't model errors — they're coordination failures. Examples: chatbots giving legally binding but hallucinated information (an airline was held liable for its bot's invented refund policy); unconstrained LLMs inventing fees or deadlines in customer service; and the silent killer — buying enterprise AI seats with no adoption plan, leaving 70%+ unused. California's design explicitly guards against these: constrained Poppy queries, human-in-the-loop review, and bundled training. The meta-lesson is that capable models fail in production when the coordination layer — access, grounding, governance, review — is missing. Build that layer before you scale.

What is MCP in AI?

MCP (Model Context Protocol) is an open standard created by Anthropic that lets AI models connect to external data sources and tools through a consistent interface — think of it as a universal adapter between LLMs and your systems. Instead of building custom integrations for each data source, you run an MCP server that exposes resources (files, databases, APIs) that any MCP-compatible model can use. For a deployment like California's, MCP is what would let Claude securely access repositories for code patching or agency data for workflows without bespoke plumbing. Its biggest strategic value: standardizing the connection layer means you can swap models without rebuilding integrations — the antidote to vendor lock-in.

About the Author

Rushil Shah

AI Systems Builder & Founder, Twarx

Rushil Shah is the founder of Twarx and an AI systems builder who has spent years designing autonomous workflows, multi-agent architectures, and AI-powered business tools. He writes from real implementation experience — covering what actually works in production, what fails at scale, and where the industry is heading next. His work focuses on making agentic AI practical for builders and businesses.

LinkedIn · Full Profile

Work with Twarx

Ready to put this to work in your business?

Twarx builds custom AI agents and automations that cut costs and win back time for your team. Book a free AI workflow audit and we will map exactly where AI fits in your operations, with no obligation.
Book your free AI workflow audit →or email hello@twarx.com

This article was originally published on Twarx. Follow for daily deep dives on AI agents and automation.

Top comments (0)