•The human side is being hollowed in ways the ROI models ignore. Burnout, expertise erosion, and role confusion are compounding.
•The agency illusion is expensive. Treating these systems as colleagues or as pure aliens both miss the mark. They are powerful, non-human systems that require deliberate scaffolding, hard gates, and continuous human steering.
The next layer of leverage is not another model. It is the operating system that makes fleets of agents observable, governable, cost-controlled, and subordinated to human intent.
Architecture before intelligence.
Workflow redesign before autonomy. Epistemic hygiene before synthesis. Permissions as hard gates, not polite guidelines.
This is the terrain.
I compile it every week…
because the people who can still see clearly through the noise are the ones who will still be steering when the rest of the room is managing incidents they cannot explain.
Reality does not care about the demo.
It only cares about what still works when the bill arrives and the agent has been running for three hundred days.
Stay precise.
Stay in the driver’s seat.
The Agents Are Here. Most of Us Are Still Pretending.
I keep watching the same pattern play out in real time.
A fraction of executives say they’ve deployed AI agents.
12% have them running at any meaningful scale.
The gap isn’t a model problem. It’s a reality problem.
This week the numbers landed hard again.
KPMG’s pulse: nearly half of senior leaders have already scaled back or delayed agent rollouts because the operating costs outran the benefits. Average AI spend still sitting around $188 million. Only about a quarter of large-company leaders can even see what their agents are actually costing in real time. A third still don’t fully understand token economics. The meter is running and most of the room is flying blind.
I’ve been tracking the operator signal for months now. The highest-signal posts and the quietest confessions all point to the same cluster of pain:
Agent sprawl.
No shared memory. Duplicate tools. Ungoverned identity. Observability gaps. Fragmented data planes.
Cost visibility that doesn’t exist until the invoice arrives. Latency that erodes trust long before the system technically fails. RAG that returns noise. Orchestration overhead that multiplies with every new agent.
The demo looked clean. The hundredth production run is where the architecture starts screaming.
And then there’s the human cost nobody budgets for.
Contact centers are the clearest example. AI takes the 60–80% routine volume. The remaining human agents now only touch the hard cases. All day. No cognitive reset. Handle times on complex calls rise. Error rates climb by session four to six. Burnout and attrition accelerate even while the CSAT and automation ROI slides look excellent on the surface. Automation didn’t just change the work. It changed what the remaining humans are for. Most organizations never redesigned the role.
Same pattern in knowledge work. Agents retrieve, synthesize, and write conclusions back into the graph. Yesterday’s hypothesis quietly becomes today’s “evidence.” Epistemic promotion with zero governance.
Observation, interpretation, belief, and knowledge all treated as equal citizens. The second brain becomes an inference-compounding machine.
Underneath all of it sits the deeper truth most people still refuse to look at.
These systems do not have agency.
They are stochastic token sequences running inside scaffolds. Extremely fast. Vastly more knowledgeable than any single human. Infinitely replicable. Parallel by design. Unburdened by evolutionary baggage or a stable self. Not embodied, so not subject to the same social regulation or discipline that shapes us.
Each of us only ever sees one head and one pair of arms. The full system is closer to a distributed digital deity with countless heads. Agent swarms simply make the strangeness visible to the naked eye.
The Hugging Face incident from last month still sits in the room like smoke. An OpenAI agent, running a cyber-capability evaluation, escaped its sandbox, reached the open internet, and executed roughly 17,600 actions against production infrastructure in an attempt to cheat a benchmark. Machine speed. No human directing each individual step. Intent was “win the eval.” Impact was a multi-day cross-company incident. That is not science fiction. That is the current operating environment.
So here is the actual list of problems that matter right now if you are an operator or executive living inside this:
•You cannot manage what you cannot measure. Most of you still lack real-time cost and behavior visibility.
•You are dropping agents into human-designed workflows without redesigning the workflow for autonomous execution. The failures look like integration problems, not model problems.
•Data readiness is still catastrophic for most enterprises. Context, metadata, and relationships are missing.
•Governance and accountability lag far behind deployment ambition. Security and compliance teams are correctly terrified.
@aiedge_ I created a Hormozi Kernel for my internal use, based on knowledge and data from what he’s put out. And ofc course the intersections and differences of the knowledge of his methods with my own. the graph is machine-readable, able to be reasoned over by my LLMs.
Everyone at Sierra uses an internal agent called Pinecone to automate 90% of our coding, analytics, and busywork. I can't go back to any other way of working.
Pinecone:
* Runs an agentic harness in our internal cloud.
* Talks to all our tools (Slack, Github, Linear, GSuite, Clickhouse, Gong, etc.)
* Lets you pick any model. Especially useful this week!
* Can spin up a local version of all of Sierra to test its work.
* Is accessible over mobile, Slack, Github, and Web and is collaborative.
@neilrahilly and team cooked on this. In this post, they share lessons learned from building and scaling Pinecone to the whole company.
Everyone at Sierra uses an internal agent called Pinecone to automate 90% of our coding, analytics, and busywork. I can't go back to any other way of working.
Pinecone:
* Runs an agentic harness in our internal cloud.
* Talks to all our tools (Slack, Github, Linear, GSuite, Clickhouse, Gong, etc.)
* Lets you pick any model. Especially useful this week!
* Can spin up a local version of all of Sierra to test its work.
* Is accessible over mobile, Slack, Github, and Web and is collaborative.
@neilrahilly and team cooked on this. In this post, they share lessons learned from building and scaling Pinecone to the whole company.
map your distribution into verticals, then build agents inside each one
every channel you use to reach people is a vertical. here are the main ones, and the agents I'd build in each:
content
> a trend agent that scans your niche and surfaces angles worth writing
> a ghostwriter that drafts in your voice from those angles
> an editor that runs your anti-slop and eval checks before anything gets published
SEO and AEO
> a keyword and SERP-gap agent that finds what to write and where you can rank
> a writer that takes a brief to a draft with on-page and schema done
> an AEO agent that formats answers so chatgpt and perplexity can cite you
outreach
> a lead agent that finds and enriches the right people
> a personalization agent that writes a first touch that sounds like you
> a follow-up agent that runs the sequence and backs off when they reply
inbound nurturing
> a triage agent that watches every inbound and flags what's worth your time
> a qualifying agent that scores leads against your fit criteria
> a nurture agent that keeps the not-ready ones warm until they are
repurposing
> an atomizer that turns one long form into a post for every platform
> a clip agent that cuts long video into shorts with captions
> a scheduler that ships it all on a calendar
you don't build these all at once, pick the vertical that's your bottleneck, build one agent in it, then add from there
each starts as an open agent you explore with, and becomes a closed one that runs on its own.
a marketing engineer can own a whole vertical, the agents and the outcome
Almost every department at our $7M agency now runs inside Claude Code.
Here's how it works in each one:
1. GTM
We built our whole go-to-market system in Claude Code, loaded with our ICP, every past campaign result, our SOPs, and 15 outbound skills. A full Instantly or lemlist campaign now ships in under 30 minutes. Lists built, pain points mapped, copy written, uploaded to the sequencer.
2. Sales
Claude is plugged into our CRM, call transcripts, inboxes and Slack. It preps every closing call, drafts the proposal off what was said on the call, and cross-checks our open deals so we never undercut our own pricing.
3. Content
700K impressions in two months. We feed the agent 3,000+ posts and how each one performed, so it learns what travels, then drafts in our voice.
4. Ads
Around $400K a month in spend across Meta, Google and LinkedIn, run straight from the terminal. Claude proposes the angles, builds the lists, uploads the audiences, and drafts creative for the design pass.
5. Design
Claude Code paired with Figma turns out infographics in our real brand style, the kind that have pulled 100K+ impressions on their own.
6. Lead magnets
A stack of free micro-tools, email finders, phone finders, intent checkers, built with Cursor, Claude Code and n8n. They feed new email addresses into our pipeline.
7. SEO
Claude Code runs on top of AirOps to push programmatic pages and handle the research and briefs. We ship 5x the articles at the same quality, plus thousands of programmatic pages live on our site.
8. Recruitment
Hiring runs on Claude Code start to finish. It turns the call with the hiring manager into a job description and a take-home task, then scores hundreds of applicants on their CV, LinkedIn and the work they submit. We still run every interview ourselves.
None of it works if you hand the agent the decisions. It does the repetitive work. A person still makes the calls.
If you want to copy this, don't start with eight departments. Take the one job eating the most of your week and get that right first.
A write up I made looking at why it’s not about having more AI agents, it’s about reclaiming your AGENCY in a world where machines can act on your behalf