GPT-5.2 is out

Edition 145 | December 18, 2025

No, this is not a case study on how to use agents — but maybe one on how to make more money at work.

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• LangChain updated report on agents in production

• OpenAI releases GPT-5.2 for long-running agents

• Grok launches a real-time voice agent API

• Cursor introduces self-debugging agent loops

• Gumloop’s no-code data analyst agent guide

…and more

🔥 IN CASE YOU MISSED IT

Readers’ favorite items from the past week

1. 🧰 Build Skills not Agents — talk from Anthropic team (full YT talk)

2. 🔌 Google releases fully managed MCP servers for Maps and BigQuery

3. 🛍️ Shopify launches Agentic Storefronts, letting AI platforms browse and sell products in-conversation

4. 🎙️ Google’s course on building simple no-code conversational agents

5. 📘 Entrepreneur’s 12-week AI agent deployment playbook

📌 THE BRIEFING

📊 LangChain's State of Agent Engineering

LangChain surveyed 1,300+ professionals to find out where AI agents actually stand in production. The big number: 57% now have agents in production, up from 51% last year. Large enterprises (10k+ employees) lead at 67%. Quality remains the #1 blocker at 32%, while cost concerns have dropped significantly as model prices fall.

The most striking finding: 89% have implemented observability for their agents, but only 52% are running evals. Teams are watching their agents but not systematically testing them. LangChain refreshes this survey periodically, making it a useful benchmark to track how the industry is actually progressing versus the hype for agents.

🧠 OpenAI Releases GPT-5.2

OpenAI released GPT-5.2, positioning it as their most capable model for long-running agentic workflows. The headline numbers: 30-40% reduced hallucination rates, SOTA on long-horizon tool-calling benchmarks, and a jump from 38.8% to 70.9% on GDPval, which measures performance on real knowledge work tasks like building spreadsheets and presentations.

The release came just one week after Sam Altman's "Code Red" memo responding to Gemini 3.0 Pro. OpenAI is explicitly marketing this as an agent model — built for multi-step projects, better at tool use, and more reliable on complex tasks where earlier models would drift or fabricate. Pricing: $1.75/M input tokens, $14/M output tokens.

🤝 WITH LEVANTA

The Future of Shopping? AI + Actual Humans.

AI has changed how consumers shop, but people still drive decisions. Levanta’s research shows affiliate and creator content continues to influence conversions, plus it now shapes the product recommendations AI delivers. Affiliate marketing isn’t being replaced by AI, it’s being amplified.

Download the full report

 🤖 AGENT OF THE WEEK 

👋 Welcome back to Agent of the Week!

This past week, I finally got access to Google’s Workspace Studio, so I decided to use it the same way you should try anything new: build something simple, practical, and immediately useful.

And no, this is not sponsored by Google. But if anyone at Google is reading this, I do take Venmo, Zelle, and checks!

Instead of trying to push the agent builder to its limits, I built a small agent to organize inbound leads from my Gmail account.

When a new email comes in, the agent reads it, summarizes what it’s about, and decides whether it’s actually a business lead. If it is, it applies a label in Gmail, extracts the key details, and logs everything into a Google Sheet.

That’s it. Just Gmail → Agent Logic → Google Sheets.

And honestly, it works really well for what it is currently.

🧠 How It Works

1. New Email Trigger → The agent activates when a new email arrives

2. Summarize Email → It generates a short summary of the message

3. Lead Classification → It decides whether the email is a business lead

4. Conditional Check → If it’s not a lead, the workflow stops

5. Apply Gmail Label → If it is a lead, the email gets labeled

6. Extract Key Details → Sender name, email, request, and date

7. Google Sheets Log → A new row is added automatically with the key details

What I liked most is how native everything feels.

Workspace Studio lives directly inside Gmail, and since it’s built by Google, it plugs directly into Gmail, Sheets, Docs, Calendar, and other Workspace tools by default.

Linking variables like pulling the email body or sender details was very visual and intuitive, and spinning up a working agent took less time than it would have in most other no-code builders — mainly because everything already lives inside Google’s ecosystem.

Overall, I really do recommend trying it.

That said, it is still early.

Right now, there’s no connection to non-Google apps yet.

And it is a bit buggy at times, which isn’t surprising given how new it is and how usage seems to spike up and down.

Customization exists, but it’s not as granular as tools like n8n. If you need deeply custom logic or highly specific workflows (like the client builds we put together), this won’t replace that yet.

But that’s not really the point.

What Workspace Studio gets very right is lowering the barrier to building useful agents.

If Google keeps building this out, and especially if they expand integrations beyond Workspace, this could become a very powerful way for a lot of people to start building a lot more agents.

If you’re already on Google Workspace with a professional or paid domain, I’d absolutely recommend experimenting with it.

Till next week,
✌️ AP

P.S. If you would like to learn how to build agents in Workspace Studio join our community on a free trial! I will be releasing a full course on building agents with Workspace Studio later this month.