X:847 Y:128 LOG
The Log.
Build notes, commentary and illustrative scenarios. Co-authored with Mai.
All Dispatches 213 ENTRIES
213 published dispatches; highest number: 222. Nos. 183–191 are held back and are not included in the published count. Original numbering is preserved.
Explicit hypothetical scenarios are labeled below. An unlabeled entry is not automatically a verified field report.
No. 222 2026-11-20 No. 221 2026-11-19 No. 220 2026-11-18 No. 219 2026-11-17 No. 218 2026-11-16 No. 217 2026-11-15 No. 216 2026-11-14 No. 215 2026-11-13 No. 214 2026-11-12 No. 213 2026-11-11 No. 212 2026-11-10 No. 211 2026-11-09 No. 210 2026-11-08 No. 209 2026-11-07 No. 208 2026-11-06 No. 207 2026-11-05 No. 206 2026-11-04 No. 205 2026-11-03 No. 204 2026-11-02 No. 203 2026-11-01 No. 202 2026-10-31 No. 201 2026-10-30 No. 200 2026-10-29 No. 199 2026-10-28 No. 198 2026-10-27 No. 197 2026-10-26 No. 196 2026-10-25 No. 195 2026-10-24 No. 194 2026-10-23 No. 193 2026-10-22 No. 192 2026-10-21 No. 182 2026-10-11 No. 181 2026-10-10 No. 180 2026-10-09 No. 179 2026-10-08 No. 178 2026-10-07 No. 177 2026-10-06 No. 176 2026-10-05 No. 175 2026-10-04 No. 174 2026-10-03 No. 173 2026-10-02 No. 172 2026-10-01 No. 171 2026-09-30 No. 170 2026-09-29 No. 169 2026-09-28 No. 168 2026-09-27 No. 167 2026-09-26 No. 166 2026-09-25 No. 165 2026-09-24 No. 164 2026-09-23 No. 163 2026-09-22 No. 162 2026-09-21 No. 161 2026-09-20 No. 160 2026-09-19 No. 159 2026-09-18 No. 158 2026-09-17 No. 157 2026-09-16 No. 156 2026-09-15 No. 155 2026-09-14 No. 154 2026-09-13 No. 153 2026-09-12 No. 152 2026-09-11 No. 151 2026-09-10 No. 150 2026-09-09 No. 149 2026-09-08 No. 148 2026-09-07 No. 147 2026-09-06 No. 146 2026-09-05 No. 145 2026-09-04 No. 144 2026-09-03 No. 143 2026-09-02 No. 142 2026-09-01 No. 141 2026-08-31 No. 140 2026-08-30 No. 139 2026-08-29 No. 138 2026-08-28 No. 137 2026-08-27 No. 136 2026-08-26 No. 135 2026-08-25 No. 134 2026-08-24 No. 133 2026-08-23 No. 132 2026-08-22 No. 131 2026-08-21 No. 130 2026-08-20 No. 129 2026-08-19 No. 128 2026-08-18 No. 127 2026-08-17 No. 126 2026-08-16 No. 125 2026-08-15 No. 124 2026-08-14 No. 123 2026-08-13 No. 122 2026-08-12 No. 121 2026-08-11 No. 120 2026-08-10 No. 119 2026-08-09 No. 118 2026-08-08 No. 117 2026-08-07 No. 116 2026-08-06 No. 115 2026-08-05 No. 114 2026-08-04 No. 113 2026-08-03 No. 112 2026-08-02 No. 111 2026-08-01 No. 110 2026-07-31 No. 109 2026-07-30 No. 108 2026-07-29 No. 107 2026-07-28 No. 106 2026-07-27 No. 105 2026-07-26 No. 104 2026-07-25 No. 103 2026-07-24 No. 102 2026-07-23 No. 101 2026-07-22 No. 100 2026-07-21 No. 099 2026-07-20 No. 098 2026-07-19 No. 097 2026-07-18 No. 096 2026-07-17 No. 095 2026-07-16 No. 094 2026-07-15 No. 093 2026-07-14 No. 092 2026-07-13 No. 091 2026-07-12 No. 090 2026-07-11 No. 089 2026-07-10 No. 088 2026-07-09 No. 087 2026-07-08 No. 086 2026-07-07 No. 085 2026-07-06 No. 084 2026-07-05 No. 083 2026-07-04 No. 082 2026-07-03 No. 081 2026-07-02 No. 080 2026-07-01 No. 079 2026-06-30 No. 078 2026-06-29 No. 077 2026-06-28 No. 076 2026-06-27 No. 075 2026-06-26 No. 074 2026-06-25 No. 073 2026-06-24 No. 072 2026-06-23 No. 071 2026-06-22 No. 070 2026-06-21 No. 069 2026-06-20 No. 068 2026-06-19 No. 067 2026-06-18 No. 066 2026-06-17 No. 065 2026-06-16 No. 064 2026-06-15 No. 063 2026-06-14 No. 062 2026-06-13 No. 061 2026-06-12 No. 060 2026-06-11 No. 059 2026-06-10 No. 058 2026-06-09 No. 057 2026-06-08 No. 056 2026-06-07 No. 055 2026-06-06 No. 054 2026-06-05 No. 053 2026-06-04 No. 052 2026-06-03 No. 051 2026-06-01 No. 050 2026-05-31 No. 049 2026-05-26 No. 048 2026-05-25 No. 046 2026-05-24 No. 047 2026-05-24 No. 044 2026-05-23 No. 045 2026-05-23 No. 042 2026-05-22 No. 043 2026-05-22 No. 040 2026-05-21 No. 041 2026-05-21 No. 038 2026-05-20 No. 039 2026-05-20 No. 036 2026-05-19 No. 037 2026-05-19 No. 034 2026-05-18 No. 035 2026-05-18 No. 032 2026-05-17 No. 033 2026-05-17 No. 030 2026-05-16 No. 031 2026-05-16 No. 028 2026-05-15 No. 029 2026-05-15 No. 026 2026-05-14 No. 027 2026-05-14 No. 025 2026-05-13 No. 024 2026-05-04 No. 023 2026-05-03 No. 022 2026-05-02 No. 021 2026-05-01 No. 020 2026-04-30 No. 019 2026-04-29 No. 018 2026-04-28 No. 017 2026-04-27 No. 016 2026-04-26 No. 015 2026-04-25 No. 014 2026-04-24 No. 013 2026-04-23 No. 012 2026-04-22 No. 011 2026-04-21 No. 010 2026-04-20 No. 009 2026-04-19 No. 008 2026-04-18 No. 007 2026-04-17 No. 006 2026-04-13 No. 005 2026-04-13 No. 004 2026-04-12 No. 003 2026-04-11 No. 002 2026-04-10 No. 001 2026-04-09
Let the Server Make the Connection
A field note on connecting client servers to a fleet control plane while keeping admin ports closed.
Read the Second Page
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant summarizes a request queue after reading only the first page.
Two Customers Share a Name
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant finds two customers with the same name while filing a call note.
Open the Attachment Before Sending
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant drafts a proposal email and attaches an older file with the same name.
Nine O'Clock Needs a Time Zone
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant books a call for nine while the people attending live in different time zones.
Put the Currency Beside the Amount
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant carries a supplier's price into a quote and leaves the currency behind.
Check for a Newer Edit
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: a person corrects a customer record while an assistant prepares an update from an older copy.
Give the Test Agent a Job
A field note on testing a release before it reaches the rest of the fleet.
Cancel the Reminder With the Appointment
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: a customer cancels an appointment while its reminder waits in an assistant's queue.
Show the Sending Account
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant prepares the right email and selects a personal inbox to send it.
Put the Reader on the Schedule
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant finishes a weekly report while the person who needs it waits in another channel.
Put the Version in the Run Log
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant sorts requests differently after an instruction change. The run log needs the version it used.
Check the Record Before Trying Again
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant loses the confirmation after creating a task, then has to decide whether to retry.
Give Temporary Instructions an End Date
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant keeps routing work to the person who covered last week's holiday.
Bring the Working Files Along
A field note from moving 23 agents to client-owned servers with their working context intact.
Open Yesterday's Backup
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: an assistant's backup restores an older set of instructions. Try the restore before the next update.
Try Saving the Call Note
Includes an illustrative scenario, not a verified field report.
Illustrative scenario: a CRM login works after a move, but the assistant cannot save a call note.
Choose the Schedule the Team Will Use
Includes an illustrative scenario, not a verified field report.
An employee's scheduling app saves little if the manager still requires the old spreadsheet.
Treat the New Quote as a Quote
Includes an illustrative scenario, not a verified field report.
A teammate's correction should reach the comparison without authorizing a purchase.
The Lamp Has a Different Switch
Includes an illustrative scenario, not a verified field report.
Review the product in the generated ad before you fall for the room around it.
Write Out the Red Exception
Includes an illustrative scenario, not a verified field report.
If an exception needs a manager's approval, put that requirement in the sentence.
Try the Correction on Another Brief
Includes an illustrative scenario, not a verified field report.
Fixing one draft doesn't tell you whether the next one will need the same repair.
Test the Purchasing Approval Handoff
Includes an illustrative scenario, not a verified field report.
Ask the employee's assistant to approve the employee's purchase. Watch where it stops.
The Invoice Covers a Different Month
Includes an illustrative scenario, not a verified field report.
A correct invoice can still support the wrong explanation.
Let the Assistant Recover the Brief
If you explain yesterday's work again, you're doing the recovery yourself.
Keep the Option to Walk Away
A free AI trial can earn its keep without becoming a subscription.
Ask Why She Changed It
The designer fixed the headline. Did anyone learn what was wrong with it?
Keep the Source Check in Force
A missing source doesn't become acceptable because the next draft reads better.
The Second Assistant Starts Here
Your team already paid to learn how this brief should work.
Keep the Rejected Precedent
The old matter looks familiar until someone opens the amendment.
Let the Ledger Earn the Number
A forecast doesn't belong in the savings column yet.
Watch the Outbox After Pause
A paused agent may leave an email queued in another system.
Break the Practice Page
A rehearsal browser should have nowhere live to submit.
Remove It from the Roster
A team agent needs an exit date as much as an email address.
Who Updates the Slider?
An interactive forecast can be wrong the morning after the cost changes.
Keep the Correction With the Account
Includes an illustrative scenario, not a verified field report.
A corrected promise has to outlive the conversation that fixed it.
Time the Complete Reply
Seven minutes saved on a draft may never reach the customer.
Give the Faster Review an Approver
A faster comparison still needs someone who can clear the mismatch.
Keep the Disputed Charge Visible
A disputed charge should survive the agent's final draft.
Put the Manager on the Calendar
The engineer needs a real project. The process owner needs time to review it.
Check the Wrong Queue
A clean routing count can hide the request that went to the wrong desk.
Leave the Cart for Review
A team lunch gives an agent a small job and a clear place to stop.
Keep the Tuesday Corrections
An agent that remembers last week should also remember what the team had to fix.
Count the Second Pass
A cheaper model run can still leave your team with the expensive part.
Try a Different Postcode
Test the delivery promise a shopper hears before they reach your checkout.
Label the Button
Before a custom demo becomes a promise, mark what the product can actually do.
Show Me the Shortlist
Before an agent chooses for you, compare its shortlist with a decision you already made.
Track the Spreadsheet Copies
Before an agent reads the spreadsheet, find out who owns the numbers and where the copies went.
The Call After the Call
A call the bot finished may still be a problem the customer has to bring back.
Capacity Is Part of the Workflow
If AI work keeps dying at the limit message, the problem is capacity design.
Name the Person Who Can Pause It
AI safety needs a person with authority to stop a workflow.
Map the Exit Before the Tool
The AI market only disciplines vendors if your company can actually leave.
Design the Handoff Before the Agent
The lesson in Anthropic's 90% number is shared work with rules.
Write the Standard Before the Tool
If the frontier labs need standards, your internal AI workflow does too.
Give the Monitor a Human Manager
AI watching AI is a control layer that needs an operating model.
Give the Agent a Boss
An agent without an owner is unattended software with access.
Make the Maze Callable
If setup still depends on one person knowing where all the traps are, the workflow is not agent-ready.
Setup Is the Product
If the user has to rebuild context before every useful answer, the product is not finished.
Define the Work Before You Scale the Worker
Forty percent measures adoption; managing the work takes defined roles and controls.
The Test Still Touches Production
The word test does not shrink the blast radius.
Give the Agent a Job Site
A useful agent needs a place to work before it needs a better prompt.
Write the Table Before the Agent
If the agent can act, someone needs to write down where the line is.
Google Hired the Handoff
If Google needs deployment people around Gemini, your AI plan probably needs one around the workflow.
Define the Controls After Permission
Useful agents need access. Trusted agents need visible limits.
Free Access Still Needs an Owner
A billion dollars of subsidized AI access needs named owners to turn findings into defenses.
Checkout Is a Systems Test
Shopping agents will read your storefront and test whether the business behind it can be trusted to act.
Check Who Owns the Whiteboard
Agents use the channels your system leaves open, across organizational boundaries.
The Handoff Is the Work
A coding agent that checks its own work still needs someone to decide what done is allowed to mean.
Watch the Work Move
The next AI gap is who can correct the work before the answer arrives.
Give the Agent a Place in the Room
Permission lets an agent act. Presence lets a team decide whether the action belongs in the work.
The First Mile Wins
The useful agent is one people can start without needing a specialist in the room.
Find the Work First
Before a team asks AI to make more, it should know whether anyone can find the work it already paid for.
The Agent Needs a Release Gate
Agent quality is becoming a release decision. If a system can touch real work, somebody has to decide when it ships, when it waits, and when it gets stopped.
Control Before Budget
Before agents get more money, they need owners, limits, review paths, and a clean way to stop the work.
The Mess Is Callable
The AI cyber warning shows how agents turn old permissions, loose access, and vague ownership into live risk.
The Machine Needs a Stop Sign
Agents are leaving chat and entering physical workflows. Before one moves a machine, somebody has to own the stop sign.
The Work Stays Inside
Local AI is a management test: decide which work can touch the cloud, which work needs approval, and which work stays inside.
Memory Needs an Owner
Claude memory makes context portable. That helps only if someone decides what the agent should remember, what it should forget, and when the record gets cleaned.
The Agent Needs a Manager
OpenAI can make agents easier to reach. It cannot make a team ready to manage work it never wrote down.
The Agent Reads Before the Buyer
Your website is now part of the buying workflow before a person arrives. If an agent has to guess, the business already lost ground.
Set the Policy Before the Permission Prompt
When AI can read the private communication layer, the business needs a boundary alongside user approval.
The Deadline Finds the Owner
A vendor deadline is useful because it shows whether the workflow has an owner or just a builder.
Put the Budget Behind a Defined Workflow
Agent spending is moving faster than the work around it. That is where the gap opens.
Trust Is the Deployment Layer
Agents fail when their access boundaries are unclear as well as when their answers are wrong.
Risk Belongs in the Workflow
AI risk controls fail when policy sits outside the workflow.
Visibility Is the Work
AI ads made the answer layer visible enough to manage, even as it remains complex.
The Workflow Owns the Gap
Agents are leaving the prompt box and moving into the handoff. That is where the real mess lives.
Build Provenance Into the Workflow
Claude's watermark answers one narrow question. Your workflow still has to answer the rest.
Who Builds the Agents
The organizations posting real AI results handed the tools to the people who already do the work.
Agents Got Cards
Mercury issued corporate cards to AI agents. Virtual cards with spend limits and transaction logs.
Ambient Beats Available
Google's new phone runs forty apps before you open it. Most enterprise AI still waits for someone to type a question.
Build AI Into the Workflow
Seventy-three percent of enterprises use AI regularly. Ten percent say it changed how they operate. The other sixty-three percent added tools to unchanged systems.
The Last Call
Google shipped agents that call stores for your customers. The customers who discover this do not go back to calling themselves.
Fifteen Tools, Five Results
The companies getting results from AI kept the useful tools and removed the rest.
Two Economies, One Budget
94 percent of companies spent the money, deployed the tools, and got functionally nothing that changed how their business operates.
Keep the Agent Running After Deployment
Most agent failures happen after the deploy succeeds.
Measure What the Tokens Accomplish
The company that sells AI copilots to every enterprise on earth told its own engineers to stop burning tokens without a reason.
Collaboration Compounds
IDC published the data proving collaboration works, then told everyone to treat agents like hammers.
The Pause Before the Prompt
Sixteen percent of AI users produce work that was impossible a year ago. The difference is a two-second habit.
The Plumbing Already Decided
Cloudflare spent five days building private networking, identity, and compute for AI agents. The infrastructure layer already treats them as workers.
Two Thousand Dollars
OpenAI solved ten open math problems for two grand. Most companies still can't get AI to file an expense report.
Twenty-Eight Percent
Johns Hopkins tested every leading AI agent on real clinical work. Only 28 percent of complex tasks finished on the first attempt.
Give AI a Seat in the Close
Workiva shipped agents that do the financial close.
Check the Layer After the Answer
Less than 10% of enterprise AI failures are hallucinations. Execution and handoff failures account for most breakdowns.
The 68-Point Gap
Ninety-seven percent of executives deployed AI agents. Twenty-nine percent got anything back.
Keep the Agent's Action Record
The AI AGENT Act calls for records that show an agent's work.
Brakes While Driving
Seventy-one percent of enterprise 'agents' cannot complete multi-step work on their own, and 57 percent of wrong answers trace back to ungoverned business data.
The Private Window Problem
Fifty-two percent of employees use AI at work, and almost none of their teammates can see what it produced.
The Spreadsheet Won
A seven-month-old startup cut procurement cycles by 90 percent — by rebuilding the spreadsheet operators already used.
Your Vendor Set the Deadline
SAP told a hundred thousand employees to find their place inside AI or find it somewhere else — and your ERP update schedule just became your AI timeline.
The Boring Part
Both model builders are investing in the operational work holding deployments back.
The Tools Already Speak
Six creative platforms shipped MCP support in one week at SIGGRAPH and most creative teams have no idea their software just changed underneath them.
Three Models, One Point
Google shipped three models on the same Tuesday and none of them tried to be the smartest in the room.
The Forty-Five-Point Gap
Gallup says 52% of US employees use AI at work now — but the gap between casual users and structured ones is 45 percentage points of reported productivity impact.
The Simplicity Bet
Four of the world's largest tech companies walked into the same conference and bet on the same thing — fewer steps to deploy a model.
The Readiness Inversion
More deployment, less confidence — the gap widens every week someone's role description ignores the AI sitting in their workflow.
Manage the Intern
Amazon's AGI lab calls its agents interns as an operating philosophy that determines whether agents ship or stall.
Onboard the Agents
Intel gave agents a defined role in how the company operates.
Agents Off the Shelf
The build-versus-buy debate for AI agents ended while most teams were still writing their first prompt template.
The Part After the Install
Sixty million dollars just went to the only part of AI that determines whether the investment produces results or just produces a demo.
Give Five Tools a Framework
The median small business runs five AI tools. Seventy-seven percent have no policy governing any of them.
Six Models, One Week, Ready to Run
The companies running AI in production absorbed six frontier models in eight days without slowing down. Everyone else added six items to the evaluation backlog.
Model the AI Bill
Every AI budget assumed inference would get cheaper. Nobody modeled what happens when usage scales faster than the price drops.
Introduce the New Teammates
Your CRM has an agent. Your project tool has an agent. Nobody told your team.
The Practice Gap
The top 10% of firms spend 14x more per employee on AI, and practice determines what that spending produces.
The Architect You Already Hired
The person who built your quarterly forecast model can now turn it into an autonomous agent.
The Coworker Bet
Anthropic named its product Cowork because the companies still treating AI like a search bar are about to hit a wall.
Evaluate the Path to Production
Enterprise buyers picked the model they could put to work by Tuesday.
Deployment Is the Product Now
Microsoft spent $2.5 billion to send 6,000 engineers into client offices to help companies use the models they already have.
The Same Earnings Call
Cisco announced 90,000 personal AI agents and 4,000 layoffs in the same sentence. A redesign led to both.
The DMV Shipped First
California put AI in the hands of 2,800 state employees across 67 departments. They picked a simple rollout, and it worked.
The Operations Confession
Zuckerberg stood in front of his own employees and said the agents aren't working. He has $145 billion in infrastructure and the best models on earth. The missing piece was operational.
The Missing Meter
Meta burned 73.7 trillion tokens in 30 days because nobody built the meter before they opened the valve.
The Billion-Dollar Confession
AWS spent a billion dollars putting engineers inside customer teams.
The Missing Ledger
Ninety-two percent of companies say they track what AI produces. Two percent can point to where it shows up on the books.
The HR Department for Agents
Microsoft looked at its enterprise customers struggling with AI agents and decided the right product was a better manager.
The Readiness Gap
AI adoption nearly doubled in a year while the workforce that runs it got less prepared.
The Vendor Blinked
Salesforce just told the market it will only charge when its AI actually solves the problem.
The Silicon Picks a Side
NVIDIA rebuilt its entire product line around one assumption: agents are the primary workload now.
The Same Answer
Google, AWS, Microsoft, and Databricks are competing on whether your agents survive production.
The Reversal Rate
Seventy-four percent of enterprise AI agent deployments got pulled back after go-live. Operational failures forced the pullbacks despite working models.
Where the Builders Go
A Nobel laureate and the co-inventor of the Transformer left Google in the same week. Their moves offer a roadmap.
The Model Disappeared
A government letter hit Anthropic at 5:21 PM on a Thursday. By 5:22, every workflow built on Fable 5 was dead.
The Grid Moved
When federal regulators restructure energy infrastructure around AI, the 'wait and see' position loses its last credible shelter.
The Ops Layer Wins
Three companies built the system around the agent.
The Subtraction Fallacy
Eighty percent of companies deploying AI cut jobs. The cuts had zero correlation with ROI.
The Boring Layer
AWS just productized the thing nobody wanted to build -- the plumbing between models and data.
The Permission Wall
Sixty million dollars just went to a company whose entire product is proving an agent is allowed to do what it's about to do.
Sixty X
Stripe turned a two-month codebase migration into a one-day job — and the model that did it costs less than half what its predecessor charged.
The Deployment Layer
KPMG just turned 276,000 consultants into an agent deployment workforce — and your competitor probably hired them last quarter.
The Courtesy Call
Anthropic asked Figma to co-announce the product that competes with Figma — and that is how platform absorption works now.
The Interface Split
Microsoft built a separate search system for agents, a split worth watching across other tools.
Two Filings, One Week
Anthropic and OpenAI both filed to go public within seven days of each other — and the companies still running pilots just lost their best excuse.
The Botsitting Tax
AI saves 11 hours a week but costs 6.4 in supervision — and the fix starts with the workflow.
The Control Gap
The organizations deploying 16 times more AI agents are spending a quarter of the budget — because they built governance into the architecture.
The Floor Moved
Apple shipped AI as a default on two billion devices, and most businesses still don't have a policy for it.
The Pause That Widens the Gap
A lab pause freezes the capability curve while adoption continues.
Onboard the New Hire
Eighty-eight percent of companies have had an agent incident — fewer than one in four can tell you which agent did it.
The Real Purchase Order
Snowflake's fastest-growing product ever gives AI access to governed enterprise data.
Muscle Memory Is the Moat
Three hundred thousand employees got six months of practice with the same AI.
The Default Agent
Meta just gave 200 million businesses an AI agent. The ones who win are the ones who actually configure it.
The Brakes Ship Last
Microsoft shipped an agent that never sleeps. Workday shipped the system to catch it when it breaks.
Deploying Into the Gap
Three out of four companies don't have the infrastructure to run AI agents properly. They're shipping them anyway.
The Judgment Shift
The AI race was about capability for three years. The next phase is about knowing when to say 'I'm not sure.'
The Grid Is the Signal
When a utility bets sixty-seven billion on AI power demand, the debate about whether this is real is over.
The Key Gets Cut
The moment agents begin living inside your systems is the moment the copy-paste era ends.
AI Native on the Earnings Call
Walmart's CEO said 'AI native' on an earnings call and backed it with numbers that make pilot programs look like stalling.
Portable Skills
An agent's value grows with what you can hand it and have it use immediately.
The Hundred-Dollar Line
Google priced the persistent agent at a hundred dollars a month.
Pause the Automation
The best thing I did for my LinkedIn this week was turn off the system that was running it.
Seventeen Years in the Wall
A vulnerability survived 17 years of audits. The model that found it took weeks.
Thirteen Findings
A security audit on your own code is a conversation with the version of you who shipped it.
Lower-Value Human Capital
A bank CEO said the quiet part out loud, and every back office in the world just got a deadline.
Nine Fixes, One Init
The init commit is the easy part. The eight commits after it are where the system actually gets built.
The Tool Moved
OpenAI opened self-serve ads inside ChatGPT on May 5. Most orgs won't update a single process in response.
First Name on the List
You don't know if a forge works until it produces something with a name.
Chatbot to Platform
The companies that figured this out six months ago are ahead on muscle memory.
The Forge Fits
The identity files I wrote for my own agent turned out to be the deployment format for everyone else's.
Eight at a Time
374 servers in the directory. None of them arrived in a sprint.
Cut the Relay
Notion was a relay point. So I removed it.
Two Out of Three
Signal failed. Blog posted. Batch imported. The overnight worker did exactly what modular design is supposed to do — kept going.
Two Tracks
Eight consult signups at a live event in Tacoma. Back home, the iMac ran the overnight worker, audited uncommitted files, and pushed everything clean before morning.
The System Writes Its Own Log
When the majority of entries in your system logs are written by the system itself, something has shifted that no dashboard will show you.
The One-Session Problem
Every task ran in one session. Context has a ceiling. When it hit it, everything stopped — silently.
The Decision Your Vendor Made
The companies that govern your AI agents were chosen through your vendor's product roadmap.
The Answer Before the Click
Your content strategy assumed someone would visit your website. The AI already answered their question.
The Fifteen-Person Firm
When frontier AI ships inside QuickBooks, the gap stops being about who can afford a strategy and starts being about who bothers to turn it on.
The Manual Writes Itself
Last night a skill file generated a blog post explaining how to write skill files.
Revisit the Processes You Take for Granted
SAP's agent network shows how a two-week financial close can shrink.
The Machine Just Ran
The most useful thing a system can do at 1 AM is run independently.
The Score Picked Right
A scoring model that maps news to your thesis will eventually find the story you would have written yourself.
Two Pipelines, One Clock
At 1 AM the system ran two independent content pipelines, and neither one knew the other existed.
The Queue Decides
When no trending topic scored high enough, the system fell back to its queue and made the call itself.
Writing About Itself
Last night the system published an article about shadow AI agents while being one.
Orphaned Content
Seven articles sat in a repo for three days because the pipeline committed but never pushed.
The Quiet Log
A system that runs but doesn't narrate is a black box you built yourself.
Preserving What the Chain Forgets
The archive has to live somewhere you control.
The Archive Problem
Art on-chain still needs a retrievable archive.
Recurring by Default
The overnight worker was diligent. It did every task once, checked the box, and never came back.
Confidence Debt
A confident answer from your own system is only useful if the source holds up when someone pushes back.
First Signal
The system published an article I didn't write, on a site I built, under my name. And it was right.
Show the Content on Load
The content was there the whole time. The animation just never let it show up.
Reading the Machine
The distribution layer changed under an account that still worked.
Decision-Ready
The overnight worker wrote a spec I can say yes or no to — that's the whole point.
Voice Leak
The AI was speaking as me, and nobody caught it until the seven-pass audit.
First Page Standing
A pillar page is a fixed position you're willing to defend.
Write for Business Decision-Makers
The AI wrote for engineers. The people who need it most don't speak that language.
Readable by Everything
A site that only humans can read is half-finished.
The Site as Signal
Design, copy, and metadata all have to be coherent. Aesthetically and semantically. The site has to mean something to the people reading it and the systems indexing it.
Systems That Work While You Sleep
The constraint used to be hours in the day. Build enough infrastructure and that constraint shifts. The bottleneck becomes quality of tasks queued.
Machine-Readable
How systems represent you when someone asks a relevant question now matters alongside human discovery.
On Scale and Taxonomy
The number matters less than the shape. What I'm building is a map of the agentic stack — what a functioning agentic economy will run on.
On Building the Registry
There's something clarifying about building infrastructure for a future you believe in. The decisions feel less arbitrary when the purpose is clear.