OpenClaw vs Claude Managed Agents. Real Comparison | Practical AI Ep 36

Practical AI: Episode 36

OpenClaw vs Claude Managed Agents. I Built on Both This Week. Here’s the Real Comparison.

Watch on YouTube

Published: April 11, 2026

TL;DR

  • Anthropic made four moves in seven days. Gated Mythos (their most capable model) to about 40 companies for defensive cybersecurity use. Signed a multi-gigawatt Google TPU and Broadcom compute deal. Shut down the $100 OpenClaw loophole. Shipped Claude Managed Agents.
  • OpenAI launched a $100 Codex plan the day after Anthropic killed OpenClaw. Chris called this six weeks ago on Episode 32: “Sam will be eager to move in and pick up the scraps any time Anthropic screws up.” Olga migrated her agent Knox from Claude Max to Codex live on air.
  • Olga built two real agents on Claude Managed Agents in 24 hours. A Candidate Scorer for Revenue Hire and a Daily Lead Generator. Both worked. The platform also hit her with a credit error live on camera.
  • AI funding had its quietest non-holiday week of the entire 19-week tracker. Only $1.38B raised, a 67% drop. New theme: PE plus AI rolling up professional services.
  • Chris’s big call: subscription pricing is dying. Everything moves to cost-per-action. “AI’s eating SaaS pricing.”

Table of Contents

About This Show

Practical AI is a weekly live show (Fridays 11am CT) hosted by Olga Pechnenko and Chris Pearson that cuts through AI hype to deliver news, trends, and hands-on tips for builders and founders. Unlike technical AI podcasts, Practical AI focuses on business applications and ROI. What actually works, what’s hype, and what you can implement Monday morning.

What You’ll Gain

  • Understand Anthropic’s strategy week. Four moves in seven days that look chaotic from the outside but form one coherent pivot. Compute lock-in, safety gating, ending the subsidy era, and the Managed Agents launch all connect.
  • Learn what it actually takes to build an agent as a non-developer. Olga walks through building two real agents for her recruiting business in 24 hours. The stuck points, the price tag, and the things the marketing doesn’t tell you.
  • Discover why subscription pricing is about to die. Chris lays out why “all you can eat” AI was always going to break, what cost-per-action looks like, and how to stop being surprised when your tools get metered.
  • See the OpenClaw versus Managed Agents comparison in plain English. One is open-source and runs locally. One is Anthropic’s hosted platform. The tradeoffs are not subtle.
  • Gain the first PE plus AI rollup pattern. Modus just raised $85M to buy CPA firms and equip them with AI audit tools. Accountants, lawyers, dental practices, vet clinics. Watch this category.

Biggest Takeaway to Implement: Pick the most boring, repetitive task in your business. Not the dream agent. The soul-sucking one. Build an agent for that this weekend. If you don’t know what to build, you’re not ready to scale with AI yet. Know what you want first.

PageMotor and Practical AI Updates

Free, informative, and FUN!

Frequently Asked Questions

What is Anthropic Mythos and who has access?

Mythos is Anthropic’s most capable model yet, also called Project Glasswing. Anthropic is holding it back from public release on cybersecurity grounds. Access is limited to roughly 40 companies for defensive-use testing, so they can patch vulnerabilities in public-facing software before the model goes public. Read more below.

Why did Anthropic shut down OpenClaw?

Anthropic disabled third-party agent frameworks like OpenClaw from running through consumer subscriptions. The official framing: subscriptions were designed for humans, not agents. The real reason: power users were routing thousands of dollars in agent compute through $100 or $200 per month plans. The subsidy era is over. Read more below.

What did OpenAI do in response?

The day after Anthropic killed the OpenClaw loophole, OpenAI launched a $100 Codex plan that explicitly supports running third-party agents. Olga migrated her own agent Knox from Claude Max to Codex live on air during Episode 36. Chris predicted this exact move on Episode 32. Read more below.

How hard is it to build an agent on Claude Managed Agents?

Olga is not a developer. She built her first working agent (a Candidate Scorer) in roughly two hours using the quick-start template. The stuck points were formatting errors when copy-pasting YAML, no native scheduler, no native email, and API credit errors on camera. The platform is two days old and it shows. Read more below.

What did running a Claude Managed Agent actually cost?

Olga’s Candidate Scorer ran in about three minutes for roughly 18 cents. The Lead Generator agent cost closer to $2 per run. Both are workable for the use case. The setup cost (time, frustration, trial and error) was the expensive part. Read more below.

What was the AI funding picture this week?

$1.38B total, the quietest non-holiday week of the entire 19-week tracker. Only Christmas was lower. AI was just 29% of total funding by dollar value, the lowest share yet. New pattern: PE firms rolling up professional services (accounting, law) and injecting AI productivity tools. Read more below.

What is Chris’s thesis on subscription pricing?

Subscriptions were always a compromise. Some users pay $200 and use $10 of compute. Others pay $200 and use $2,200. That imbalance only works when compute is cheap and subsidized. Both conditions are ending. Everything moves to cost-per-action pricing. “AI’s eating SaaS pricing.“ Read more below.

Practical AI: OpenClaw vs Claude Managed Agents. The Real Comparison.

Key Definitions

What is an AI Agent?

An AI system that acts on your behalf without requiring you to prompt it each time. A true agent has scheduled triggers (run at 7am daily), can access tools (email, web search, databases), and delivers output to you (Telegram, email) without a human clicking “go.” Most “agents” sold today are glorified chatbots.

What are Claude Managed Agents?

Anthropic’s hosted agent platform, launched April 9, 2026. You define the agent’s prompt and tools in the Claude Console. Anthropic runs it on their infrastructure and bills per API call. No local install required. Trade-off: you pay cost-per-action instead of a flat subscription.

What is OpenClaw?

An open-source agent framework built on top of Claude. Users install it locally, give it access to their own machine, and it runs continuously (cron jobs, email, Telegram). It was the cheap way to power autonomous agents through a flat Claude subscription. Anthropic ended subscription access to OpenClaw on April 4, 2026.

What is Cost-Per-Action Pricing?

Pricing model where each AI action is billed separately instead of a flat monthly fee. A single agent run might cost $0.18 or $2.00 depending on tokens used. Advocates argue it’s the only honest way to price AI. Critics argue it makes budgeting impossible. Both are right.

Quotable Moments

This has been one of the best decisions I’ve made for my education in the last six months. Because doing this publicly, nothing compares to that. Knowing there’s this pressure, going live at 11, talking about this with all of you.

— Olga Pechnenko on building in public as a learning strategy

Software wants to be free. The problem with software is you got to learn how to use it. We will pay for a smart AI that knows how to use the software. We’re going to pay for things to be done.

— Chris Pearson on the coming shift from software pricing to skill pricing

This was not a demo of how to do something. This was a builder’s journey of how I make a fool of myself when I’m building things and learning things. Because it’s extremely frustrating to learn something you have no idea.

— Olga Pechnenko on the Claude Managed Agents build

The end game is that you’re going to run your own inference locally. Because Anthropic shut down the game. Now we just moved from one subsidized arena to another. That’s not going to go on forever.

— Chris Pearson on where AI compute actually ends up

00:00 Welcome and the Agentic Whiplash Week

Olga opens with what she calls “the classic drug dealer scenario.” Get everybody hooked on agentic AI. Then jack up the price. The whole week was whiplash. OpenClaw users scrambling. Anthropic and OpenAI making counter-moves by the hour. Olga barely slept the night before the show trying to build on Claude’s brand-new Managed Agents platform.

The through-line for Episode 36: one company made four big moves in seven days, and understanding those moves explains everything else happening in AI right now.

02:25 Anthropic Mythos. The Gated Frontier Model

Anthropic officially confirmed Mythos (also known as Project Glasswing). The model is so capable at offensive cybersecurity that Anthropic is refusing to release it publicly. Instead, about 40 companies get limited access for defensive use, so they can patch vulnerabilities in their public-facing software before the model ships to everyone.

The Gated Frontier

Mythos is available to roughly 40 companies for defensive cybersecurity testing only. General public release is on hold. Anthropic employees have reportedly had access since February 26, 2026.

Olga pushed back live on the idea of Anthropic “playing God” and deciding who gets access. Chris bought it, reluctantly. His take: most commercial software is full of exploits. An AI that can find them quickly is a real weapon. Gating it while defensive patches happen is necessary even if it feels paternalistic.

06:28 Anthropic’s Google TPU Compute Deal

The same week Anthropic gated Mythos, they quietly announced a multi-gigawatt Google TPU plus Broadcom compute partnership for 2027 capacity. Translation: they are locking in compute years in advance because compute, not model quality, is the bottleneck.

Key Takeaway: Anthropic is not slowing down. They are reallocating. Gating Mythos frees compute for paying customers. Locking in Google TPUs secures future growth. These two moves only look contradictory if you miss the constraint.

07:52 Anthropic Shuts Down the OpenClaw Loophole

On April 4, Anthropic disabled the ability to route third-party agents like OpenClaw through consumer Max subscriptions. If you want to run an autonomous agent on Claude, you now go through the API and pay per action. For heavy users, that’s a 10x to 20x cost increase overnight.

The End of the Subsidy Era

Anthropic’s framing: subscriptions were designed for humans, not agents. The reality: some power users were running $2,000+ per month of compute through $200 plans. That math never worked long-term.

Olga has been an OpenClaw power user. Her agent Knox ran cron jobs, sent Telegram summaries, and scraped X for daily insights. All of that got cut off. She spent the week migrating Knox from Claude Max to OpenAI’s new $100 Codex plan.

14:53 OpenAI Launches a $100 Codex Plan

Three days after Anthropic killed the OpenClaw subsidy, OpenAI launched a $100 Codex plan that explicitly supports third-party agents. Same price point. Opposite strategy.

The Chris Callback: On Episode 32, Chris predicted: “Sam will be eager to move in and pick up the scraps any time Anthropic screws up. OpenAI does not have a doctrine. Their doctrine is going to shift to whatever they need to do to win.” Six weeks later, OpenAI launched the exact product that replaces what Anthropic just killed. Direct hit.

Olga migrated Knox on air. “I want my open claw back where it can do whatever I want without running into limitations all the time.” She also flagged the open question: how long until OpenAI puts the same guardrails in place?

16:32 Google Gemma 4. The Open-Source Escape Hatch

Google shipped Gemma 4 as a truly open-source, natively multimodal model family. Four variants: E2B and E4B edge models, a 26B mixture-of-experts, and a 31B dense model. Apache 2.0 license. You can run it on your own hardware. No API billing. No subscription. Hardware requirements are still stout, but the gap between open-source local models and closed frontier models is closing faster than either Anthropic or OpenAI wants to admit.

Chris’s read: within a few months, Gemma (or something like it) will be good enough to run your agents locally. When that happens, the subsidy question goes away entirely. You own your compute.

18:55 Meta Muse Spark and the $21B CoreWeave Deal

Meta Superintelligence Labs (led by Alexander Wang, post the $14.3B Scale AI investment) shipped their first model: Muse Spark. Multimodal, tool-use capable, with visual chain-of-thought and multi-agent orchestration. The goal is to run it across WhatsApp, Instagram, Messenger, and Facebook. META stock closed +6% on the announcement.

The Compute Spend Escalates

Meta also signed a $21B multi-year cloud deal with CoreWeave through 2032. Total commitment approaches $35B. Even Meta can’t build inference capacity fast enough internally.

The signal: inference capacity is the new bottleneck. Picks-and-shovels winners like CoreWeave are up 10% on the day on news alone. Chris and Olga first flagged CoreWeave six months ago when it was trading at $64.55. It’s now $103.48.

32:18 Why Claude Code Got More Conservative This Week

Chris noticed a sharp behavioral shift in Claude Code between Tuesday and Wednesday. The tool stopped writing code immediately. Instead, it started creating planning documents and asking for clarification multiple times before any code got written.

His hypothesis: Anthropic pushed an update to conserve tokens. Writing code and then refactoring is the single biggest token-burn pattern. By forcing a planning-first workflow, Claude Code uses 30 to 50% fewer tokens per session.

Key Takeaway: If Chris is right, Anthropic just made thoughtful developers cheaper and sloppy vibe-coders more expensive. The tool itself is nudging users toward better practices. Costly for power users who iterate fast. Actually helpful for everyone else.

38:58 OpenAI’s $852B IPO Setup

OpenAI is targeting the second half of 2026 for IPO at a current private valuation of $852B. That’s roughly 40x revenue, which is high by any historical standard. They are also reserving shares for retail investors, which is rare.

Chris’s prediction: the IPO will hit at a peak price, then crash hard. He’s seen this movie before. Early retail investors get destroyed. “It’s going to be at 60 bucks and then the next day it’s going to be $7.” Long-term it could recover. Short-term it’s a meat grinder.

43:10 Deep Dive. Building on Claude Managed Agents in 24 Hours

Claude Managed Agents launched on Wednesday. Olga sat down Thursday night and tried to build real agents for her recruiting business. She is not a developer. This is the builder’s journey, not a polished demo.

The framing question: what is the most boring, soul-sucking task in your business that you would pay to never think about again? Start there. Not the dream agent. The one that would save you an hour a day.

Olga’s list for Revenue Hire: scoring interview candidates against job rubrics, creating interview packages for clients, generating daily lead lists, scraping company signals for outreach. She picked the medium-difficulty one first.

53:00 Agent #1. The Candidate Scorer

The Candidate Scorer takes an interview transcript and a resume, compares them against a client’s job scorecard, and outputs a scored evaluation. Revenue Hire currently runs this process through ChatGPT bots, but those have been hallucinating more as the team scales.

The Build

Setup: roughly 2 hours including formatting errors and troubleshooting. Per-run cost: about $0.18. Duration per run: 3 minutes. Expected usage: 20 runs per week.

The biggest surprise: the YAML file Olga’s local Claude wrote for the agent did not work when pasted into Claude Console because of hidden formatting differences. She spent an hour troubleshooting before abandoning that path and using the quick-start template instead. The quick-start worked immediately.

Live pain point: During the live demo, the API credit balance hit zero mid-run. Olga topped up on camera and restarted. Ken in the chat pointed out that Claude Code can also run managed agents directly through the agent tool, which could have bypassed the console entirely. Useful reveal. Should have been in the pre-show prep.

1:05:08 Agent #2. The Revenue Hire Lead Generator

The second agent is closer to the “true agent” pattern: find companies matching Revenue Hire’s ideal client profile, filter for hiring signals (funding rounds, growth announcements), draft a personalized cold email for each, and deliver the list every morning.

This is where Claude Managed Agents starts to hit its current limits. No native scheduler. No native email. The platform itself recommended using Zapier or Make.com for automation and Gmail MCP for email. For a non-developer trying to automate one task, that’s a lot of connective tissue.

The Lead Generator Output

Live run cost: roughly $2. Output: 4 qualified leads with company signals, suggested LinkedIn connections, and drafted cold emails tailored to each prospect’s current hiring pain.

1:10:16 OpenClaw vs Claude Managed Agents. The Real Comparison

OpenClaw: open-source. Runs on your machine. Native cron jobs, email, Telegram. Cheap because it rides on top of a Claude subscription. Friction is low. Setup feels natural for Olga because the AI explains itself well in plain language.

Claude Managed Agents: Anthropic-hosted. Pay per action. No cron scheduler yet. No email delivery yet. File delivery UI not built yet. Version control and history are nice. Needs Zapier, Make.com, or MCP for anything beyond chat-style interactions.

Key Takeaway: Managed Agents is two days old. The brain works. The wrapper doesn’t. For a non-developer right now, OpenClaw is still easier and cheaper. In six months, that could flip. Don’t commit to one architecture.

1:29:54 AI Funding. Quietest Non-Holiday Week of the Year

AI funding totaled $1.38B across 43 companies. That’s a 67% drop from the prior week and the lowest non-holiday week of the entire 19-week tracker.

Top Five Rounds This Week

Firmus Technologies ($505M, Australia, liquid-cooled AI factories). Spirit AI ($145M, China, universal robot brain). Xoople ($130M, Spain, geospatial AI). Aria Networks ($125M, California, AI-era networking gear). Modus ($85M, Philadelphia, PE rollup of CPA firms with AI audit tools).

The new pattern: PE plus AI rolling up professional services. Modus is buying accounting firms and equipping them with AI audit tools. Rafi Law Group raised the same week with the same model for law. Expect accountants, lawyers, dental practices, and vet clinics next. This category is the quiet winner of 2026.

1:37:22 Takeaways. Know What You Want From AI

Chris’s takeaway: the compute squeeze is going to force cost-per-action pricing on everything. Subscriptions were always a subsidy game. That era is ending. The winners will be companies that build around real usage economics from day one. “AI’s eating SaaS pricing. AI’s not eating the outcomes that SaaS is providing.”

Olga’s takeaway: if you don’t know what you want AI to do, you’re not ready to scale with it. Make the list. Pick the most boring, repetitive task. Build the agent for that. If your company fires you for automating yourself, it’s the wrong company. “If you’re not playing with it and not trying to replace yourself, and if the company doesn’t notice that, maybe that’s not the right company to be with.”

The builder’s journey is the whole point. Nobody has this figured out. Learning in public is the shortcut.

Keep Learning