Also: 55% of jobs reshaped, power over GPUs, a thousand DIY weekend platforms, and mood-tracking the robot
MCP Servers for, and of, knowledge workers
This is an excerpt from last week’s Tanzu Catsup. David and I have been talking a lot about the AI harness. We of course sprinkle in talk about how you would enterprise-y it up, esp. with security.
My key insight and angle on all this is thinking about how all the AI harness stuff applies to NOT programmers, to knowledge workers. That’w where the huge benefits will accrue, the huge amounts of compute will be driven, and where the big changes will happen, and need to.
More excerpts from that same episode: thinking of swarms of agents and having to click accept all the time really kills the vibes.
Related to your interests
- The Real Constraint on Enterprise AI isn’t GPUs; It’s Power - “When AI workloads are deployed in isolation and utilization is around 50-60%, we often see significant waste in server capacity as well as improperly provisioned power distribution for the expected peak load that didn’t materialize.”
- Unified Self-Service Consumption for Modern Workloads - “Improving integration between tools and systems, which enables a unified platform, is the most desired organizational technology improvement, cited by 41% of 250 respondents in 451 Research’s Voice of the Enterprise: Workforce Productivity & Collaboration 2025 survey.” // “43% of 273 respondents project that the preferred IT environment for generative AI workload deployment in two years will be private cloud (on-premises, colocation, or hosted) as opposed to public cloud and SaaS alternatives.”
- Sounds like someone got a 25% discount - “Amazon.com Inc. plans to invest up to $25 billion in Anthropic PBC and provide the company with a significant amount of additional computing capacity…. Anthropic, for its part, has committed to spending over $100 billion on Amazon Web Services over the next decade.”
- Although, that also looks like another "maybe we’ll spend that much" situation: “As part of an expanded pact, the e-commerce and cloud giant is poised to invest an additional $5 billion in the Claude developer now and up to $20 billion more in the future. Anthropic, for its part, aims to spend more than $100 billion on Amazon Web Services tech over the next decade.”
- BCG: 50-55% of Jobs Reshaped by AI in 2-3 Years, Only 10-15% Displaced - Entry-Level Is the Real Casualty - The tools you use drive how you work. When you have a new tool, if you don’t change how you work, you don’t get the full benefits of the tool. Incremental innovation vs. innovation innovation.
- Early AI adopters stumble across the horror of vendor lock-in. Here we go again? - “of the 2/3 who have actually tried a [AI service, etc.] migration , only 42% report a smooth transition and 58% say it failed or blew way past estimates.” // You only call it “lock-in” when it stops working and/or you don’t like the pricing. Otherwise you call it “productive” and “strategic.” // As someone else put it: “what you pay for is work done."
- The AI engineering stack we built internally — on the platform we ship - Cloudflares AI coding harness, at an enterprise-grade level. // This is like a reference architecture for everything needed to make an enterprise setup - it lacks the depth of governance and compliances that a bank would want. // Also, this launches thousands of DIY platforms projects where a developer said ‘I could do that in a weekend,’ brining all the downsides of DIY platforms to a new era of enterprise computering. // Looks cool!
- Agent Harness Engineering - “A decent model with a great harness beats a great model with a bad harness.” // An overview of what AI harnesses look like and some tips on using them. Plus, this whole notion is evolving and changes as the models get better, or, at least change.
- Designing Memory for AI Agents: Inside Linkedin’s Cognitive Memory Agent - InfoQ - 🤖 “The broader signal: production AI is defined less by the model and more by the memory, context management, and infrastructure around it. CMA is LinkedIn’s bet on externalised memory as a horizontal platform for adaptive, personalised agentic systems at scale.”
- How The Access Group Scaled AI-Driven Development with Platform Engineering - “That’s the shift that matters. When you provide the right level of abstraction, you don’t just make systems easier to use; you make them easier to own. Engineers can take responsibility for the parts of the platform that matter to them, evolve those capabilities independently, and own their flow of value.” // This is what kicked off my tinker-talk.
- This seems pretty good, actually. We’ve put the robot in talky-talk mode in the middle of the table at dinner several times and has good results re: family quality time. I mean, at one point in history, Socrates was like “next thing you’re going to tell me is that families (well, of course not the women) are going to sit down and talk about what’s in a book. That will ruin society. Also, does anyone really know what ‘dinner’ is?”
Wastebook
- “Every major enterprise-category vendor is shipping the same three-line marketing - ‘agents,’ ‘guardrails,’ ‘control plane’ - and the thing to watch is which one ends up with a durable product-shape in eighteen months.” - Claude summarizes recent enterprise AI marketing.
- Nevermind mood journaling: I bet you track my mood over time by looking at my responses to the robot in AI chats. Happy: “nice, homie!” “Love it, home skillet!” Moody: “the fuck is wrong with you? You did this perfectly yesterday and last week.” “Stop asking shit-brained questions and just do the fucking shit I fucking asked for.” (To be fair…to me…those moodiness e.g.’s were a but arch. I’m not quite that toxic of a user.)
ICYMI
- Let them tinker - hacking developer resistance to sound enterprise architecture and platforms - Instead of locking the enterprise platform down, let developers customize and experiment - that’s how you get adoption without losing the org benefits of a shared platform.
- Macro-economic headwinds vs ‘The Terrace-in-April’ - Europe’s comfortable quality of life is real, but the external conditions propping it up are shifting, so growth isn’t optional if you want to keep the terrace in April.
- Giving agentic AI a sandbox that won’t delete everything - Tanzu Catsup - David and I discuss some recent attempts to put more controls over free-wheeling agentic AI, here, how Claude code uses the bash shell and accesses data sources. Also: some experiments with using Tanzu Platform (Cloud Foundry) as a backend for shell commands. David also goes over a better option - an MCP Server - and how he’s been architecting swarms of agents patterns.
Logoff
I’ve been struggling of late to find relevance in what I’m interested in, do, and work on. To worship at the alter of the cult of businesses: yes, at the end of the day I can feel the bulk of activity I’ve done - there’s definitely very little Terrace-in-April in my life. But, is all that work relevant to where I was in the industry, what I was doing, and the “work” I felt I was up to several years ago when I was really pumping the stuff out? I type this no to whine (well, sure - yes) but more to find a tool to judge if this is overall “ok” and “expected.”
My pal Seroter loves linking to the occasional career and enterprise-cog-mental-therapy piece in HBR. There are two varieties: (1) take moments like these to chill, burn off burn-out kindle, and arrange your enterprise sock drawer to be ready for whatever comes next, or, (2) take this a a signal to burn the boats with a fucking flame-thrower and thrust yourself into the great unknown of the “now more than ever” enterprise-scape.
I have to say, I’m not into boats, so I don’t think I even have any.
Want to subscribe to this newsletter and get it in your email? Do that here. You’ll just get this type of link and post round-up, not everything posted on the weblog.