Building an Anthropic Managed Agent for 6 AI Products — Live Stream Recap

Anthropic dropped something big this week: Managed Agents — their new framework for building production-ready AI agents that can handle complex, multi-step tasks.

I went live to build one from scratch. Not a tutorial. Not a demo of someone else's code. Just me, Claude Code, and the problem of running 6 AI products as a solo founder.

What I Built

The goal was simple: create a managed agent that could help orchestrate tasks across my product suite:

• ActorLab — AI tools for actors (19 tools, 159+ scenes)
• CastAlert — iOS casting notifications
• Prelithic — 3D pyramid reconstruction (530K+ blocks)
• noui.bot — Agent infrastructure
• TrashAlert — San Diego trash pickup schedules
• LIMS BOX — Lab information management

That's a lot of context switching for one person. The managed agent needed to understand all of them.

The Architecture

I'm already running a 24/7 ops agent (Daisy) on a Mac Mini. The managed agent needed to fit into this existing system without breaking what works.

Key decisions:

• Opus 4.6 for orchestration — the brain that delegates


• Haiku for routine tasks — cost optimization that matters


• Flat $200/mo cost model — no surprise bills, predictable runway

The managed agent pattern lets you define clear handoff boundaries. When one agent completes its scope, it hands off cleanly to the next one. No context bleed. No confusion about whose job is what.

What Surprised Me

The handoff pattern was the revelation. Managed Agents aren't just about delegation — they're about graceful delegation. When Claude knows it's handing off to another agent, it prepares context differently.

This matters when you're running multiple products. Context bleed between projects was my biggest problem before. Building ActorLab features while Prelithic monitoring ran in the background meant constant mental overhead. Managed Agents helped compartmentalize.

The other surprise: how much you can accomplish with thoughtful model routing. Opus for thinking, Haiku for doing. The cost difference is 10x, and for most routine operations, Haiku is plenty smart.

The Numbers

After running this setup for a day:

• Total API cost: ~$12 (well under the $200/mo budget)


• Tasks delegated: 47 across 6 products


• Human intervention required: 3 times (all approval gates I designed)

That's the goal. Build systems that run themselves, with humans in the loop only where it matters.

Watch the Full Stream

If you want to see the actual build process — mistakes included — here's the replay:

📺 Watch: Anthropic Managed Agents — Building a Claude Agent for 6 AI Products

The stream is about an hour long. I didn't edit out the debugging or the moments where I had to think. Real builds are messy. That's the point.

What's Next

I'm spending the next week stress-testing this setup across all 6 products. The real question isn't whether managed agents work in a demo — it's whether they hold up under production load with real users and real edge cases.

I'll be posting updates:

• X thread breaking down the architecture (Day 2)


• Community posts in VCA and AI Hustle with lessons learned (Day 4)


• Short clips from the stream with the key moments

Follow along: @HudBeer on X.

---

This is part of the "Behind the Build" series where I document how ActorLab and the rest of the TombStone Dash portfolio gets built — one commit at a time.