Agent Sprawl Ends Here: Building an Agentic OS with Claude Code (on AWS)

Gartner expects 40% of agentic AI projects to be cancelled by 2027. The ones dying aren’t failing at the model - they’re failing at the boundary: agents scattered across laptops, shared credentials nobody can revoke, and no way to answer “which agent touched that data?” Mark Kashef built an Agentic OS that answers those questions, and runs it inside a customer’s own cloud: identity, guardrails, and an audit trail that make it safe to hand real work to a team of agents. Not another dev tool, not a “second brain” - the operating layer a company runs on, a Business OS sliced to your team. He’ll cover what to adopt versus build, and what actually survives an architecture review board.

The arc. Part 1 showed you could build multi-agent systems from scratch. Part 2 showed the model call is the easy part - the real work is everything around it. Part 3 asks what’s still worth building yourself, now that all of that ships as a governed layer on the cloud you already run. The thin slice that IS your business, not the platform underneath it. That’s the serverless mindset: build just enough to serve the business, never technology for its own sake.

What you’ll take away

• Why laptop-scale agent stacks (OpenClaw, Hermes) break the moment they meet a real company.
• The five layers of an agentic OS, and why a dashboard over broken plumbing is theater.
• The cockpit: start with your own “what do I do next,” then scale it to a governed slice for one department - legal, finance, HR - with a human at the controls.
• What to build: the platform already carries the heavy lifting (Bedrock, Vertex AI, AI Foundry, Databricks), so you build only the thin slice that IS your business.

Who this is for
Enterprise IT architects, platform engineers, and developers bringing agentic AI into an organization with real security, compliance, and audit expectations - whether you are evaluating, piloting, or defending a design to a review board.

Format: 30-minute talk, then 30 minutes of live Q&A. Bring your hard questions - this crowd always does.

Speaker: Mark Kashef is the founder of Prompt Advisers and the instructor behind Early AI-dopters, a 1,300-member community for practitioners building with Claude Code and Codex. He is an O’Reilly Media instructor, a former AI and analytics professor, holds a Master’s in AI, and is ex-Amazon. His YouTube channel (80k+ subscribers) is one of the most-followed practitioner sources on agentic coding. Mark returns to Serverless Toronto for a full solo session after his short segment at our February anniversary.

Homework (optional, but you’ll get more out of the Q&A)
Mark builds this in public. The best primer is his walkthrough of the five layers, which is the spine of this talk. The rest of the series: Agentic OS on Mark’s channel.

Serverless Toronto has been bridging the gap between IT and business needs since 2018. Past sessions: youtube.serverlesstoronto.org