Georgiy Tarasov, AI Product Engineer at PostHog, shares what PostHog learned from seven experiments building interfaces for AI agents — and why MCP isn’t always the right answer.

As customers started asking for “all PostHog AI features in my agent,” PostHog had to rethink how its product should work outside the browser: inside Claude Code, Cursor, Codex, CLIs, MCP clients, and local agent workflows.

In this talk, Georgiy walks through PostHog’s experiments with handwritten agent tools, traditional MCP, CLI-first interfaces, MCP with a CLI-like shape, code execution, dynamic toolsets, and SQL-based retrieval. He compares the trade-offs across developer experience, context bloat, tool discovery, token efficiency, latency, eval results, and real customer usage.

The core lesson: your new user might not be a human in a browser tab — it might be your customer’s AI agent.

👉 Subscribe to Itnig for more conversations about real business, startups and brands.

🎙️ Want to join the Itnig podcast or sponsor one of our episodes?

Appear on the podcast: https://tally.so/r/wo1Poe

Sponsor the podcast: https://tally.so/r/3EERLN

ABOUT ITNIG: 🐦 X - ⁠https://x.com/itnig

💡 LinkedIn -⁠ https://linkedin.com/company/itnig

📸 Instagram - ⁠https://instagram.com/itnig

💌 Newsletter - ⁠https://itnig.net/newsletter/⁠

🌐 Web -⁠ https://itnig.net/

LISTEN TO OUR PODCAST ON:

🔊 Spotify: ⁠http://bit.ly/itnigspotify⁠

🎙️ Apple Podcast: ⁠http://bit.ly/itnigapple

⁠00:00:00 Intro & welcome — AI Builders BCN first edition

00:01:04 Georgiy introduces himself & PostHog

00:02:14 The challenge: shipping AI agents on a complex product

00:04:37 First MCP experiments & why they didn't scale

00:05:26 Token optimization problems & early lessons

00:07:08 Rethinking the approach: sandboxing & unified interfaces

00:08:13 Deep dive: how MCP prompts & instructions really work

00:10:26 Optimizing MCP tools for different model providers

00:12:55 Experiment #1 — grouping tools by intent00:14:15 Switching to a resource-based approach

00:18:24 Advantages & limitations of the code execution approach

00:22:30 MCP with native code execution — Georgiy's favourite

00:25:15 How GitHub optimizes their MCP (grouping by intent)

00:25:43 Agent skills: what they are and why they matter

00:26:43 Building skills with regex, access controls & data

00:27:52 Benchmarking: testing with real vs synthetic data

00:29:38 Results & what actually worked in production

00:31:35 Key takeaways & closing thoughts

00:31:43 Q&A

00:36:29 End

Recorded at AI Builders Barcelona.

Speaker: Georgiy Tarasov, AI Product Engineer at PostHog

Topics: MCP, AI agents, Claude Code, Cursor, Codex, CLI, codegen, agent interfaces, developer tools, PostHog, AI engineering