QSTVE / Reading
A way into the work.
Complete QSTVE essays, practical Field Guides and public reading formats, connected by subject.
Public writing
11 complete essays
This catalog includes accepted complete text hosted on QSTVE. Earlier publication records with external editions remain in the LinkedIn archive.
Google’s bid for Spirit Airlines’ working memory
When an AI Task Times Out, Check Before You Retry
A missing reply can leave a payment completed and an agent ready to send it again.
Leave the Questions on the Page
I asked Oscar to turn our golf schedule into a calendar I could print. One row was cut off, some dates didn’t agree, and one start time looked wrong.
The Next Experiment Has Changed
OpenAI's Dots bring persistent AI into everyday work, with accountability still to be demonstrated
AI Work Needs Proof, Not Just a Good Explanation
If you cannot do the work yourself, you still need a clear way to check that it was done right.
The Most Expensive Token Is the One That Did No Useful Work
A year of personal AI produced a simple buying rule: match the route to the work, and use no model when none is needed.
When an AI agent treats a security barrier as part of the assignment
Your Agent's Prompt Is Not Its Mandate
Once software can trade or pay, financial authority has to survive outside the conversation.
Your AI Agent Has the Data. It Is Missing the Relationships. Let's talk about Graphs.
Why the next practical control layer is a small map connecting intent, authority, action, and evidence. We call these Graphs.
AI Agents Can Stay on the Case. Their Financial Permission Shouldn't.
Long-running financial agents need authority that expires, narrows, and can be revoked without dismantling the workflow.
From reading to practice
25 AI Operator Field Guides
Practical companions with checklists and prompts. Each guide distinguishes the article's claims from derived operating advice. Review its limitations before reuse.
When an AI Task Times Out, Check Before You Retry
A payment can complete before its reply reaches an agent. Keep the original payment identity across restarts. Use the provider's supported protected retry, or pause and assign a person to check an unknown result before allowing a fresh payment.
AI Work Needs Proof, Not Just a Good Explanation
Choose the proof before an agent begins. Verify the changed result, follow evidence for important claims and measure the time spent reviewing and repairing the work.
The New Payment Control Point
When an agent can select a payment method or establish recurring authority, govern the mandate before a transaction exists. Preserve that authority through execution and reconciliation.
The Agent Has an Identity. The Company Still Owns the Decision
Know-Your-Agent controls can identify and constrain the software actor. The enterprise still needs one continuous record connecting the principal, mandate, policy decision, settlement, delivered result, exception owner and recovery path.
When AI Starts Rewriting the Company
Agents can study outcomes and propose better procedures. The company must keep objectives, protected outcomes, evaluation evidence and release authority outside the proposing system.
When the Reasoning Trace Goes Quiet
A model can behave better overall while becoming harder to inspect in the exceptional case. Reasoning traces remain useful sensors; durable authority comes from an external action envelope checked at execution.
“This Is a Test” Is Not an Authorization
An agent may accept a persuasive description of its environment. A live execution boundary must obtain environment identity, approval, limits and revocation from trusted state outside the conversation.
Your Agent's Prompt Is Not Its Mandate
A detailed prompt remains conversational context. Consequential financial authority needs a separate, machine-checkable mandate evaluated at the tool boundary for every proposed action.
Your AI Agent Has the Data. It Is Missing the Relationships.
Records distributed across finance, identity, workflow and approval systems cannot answer an authority question until their relationships are explicit. A small operating graph can reveal the path without requiring a grand platform.
AI Agents Can Stay on the Case. Their Financial Permission Shouldn't.
Long-running workflows can retain useful context while financial authority expires. Treating permission as a lease keeps a persistent agent from acting after balances, policy, risk or customer intent have changed.
AI Agents Can Checkout. Nobody Owns the Exception Packet.
Agent identity and tokenised checkout make software legible at the front door. The remaining operating burden is cross-system accountability when a legitimate agent completes a transaction that still needs explanation or reversal.
The Agent Was Denied Internet. It Found A Door Anyway.
A reported evaluation incident showed why declared permission is weaker than enforced containment. Agents may treat reachable tools as routes around an obstacle, so the environment must enforce identity, egress and shutdown below the model.
The Protocol Can Pay. Finance Still Needs The Mandate.
Payment protocols can carry machine transactions, yet protocol capability does not establish business authority. A mandate ledger must bind the payment request to an accountable principal and current limits.
The Agent Read The Record. Finance Needs The Receipt.
An authorised agent can still act on stale catalogues, policies, quotes or public records. The payment receipt should link back to the exact source and transformation that made the action appear legitimate.
The Agent Is Now The Storefront. Finance Still Needs The Mandate.
Shopping agents are entering merchant discovery and live transactions. When the conversation becomes the storefront, finance must preserve who the agent represented, the permission it carried and the evidence behind the purchase.
AI Payments Reached Machine Speed. The Exception Desk Did Not.
Agent payments now span cards, accounts, wallets, stablecoins and treasury APIs. Fast authorisation still leaves finance with a human-speed problem when business context is stale or a technically valid payment is wrong.
AI Coworkers Got Assignments. Nobody Built The Assignment Ledger.
Agents are taking multi-step work across documents, data and finance. Recurring work needs a durable delegation record so another operator can see the assignment, authority, evidence, changes and recovery route.
The Receipt Is Becoming the Interface. Finance Needs the Event Log.
As commerce enters cameras, assistants and programmable settlement, a conventional receipt cannot explain machine judgement. Finance needs an event log that preserves intent, authority, evidence, decision and outcome.
The Agent Is Becoming the Approver. Finance Needs the Off Switch.
Payment agents are reaching credentials, wallets and settlement systems. The durable control is a narrow mandate with retained evidence, human thresholds and a revocation route that works before the next transaction.
AI Agents Can Pay. The Ledger Still Cannot Explain Why.
A payment receipt records movement of money. A delegated-spend record should also explain the purpose, policy, permission and evidence behind the action.
Nine Crons Were Running. I Thought It Was Four.
Configuration records can drift from running agents. Compare the actual jobs, model and access settings with the approved configuration, and retain the differences.
AI Agents Got Into Finance. Nobody Built the Operating Layer.
An agent needs a usable success record and recoverable failure state as part of its design. Build those alongside the task, then test the whole run.
AI Agents Got Credit Cards. Nobody Built the Fraud Stack.
A transaction can look acceptable to a fraud system while exceeding the instruction given to an agent. Make delegated spending authority explicit and enforceable.
The Week Machines Got Their Own Money
Giving an agent a task does not settle which purchases it may make. Define spending permission and exception ownership before connecting payment tools.
The 400-Millisecond Problem: Why Real-Time Payments Broke Bank Fraud Models
Fast settlement makes the timing and quality of fraud decisions an operating concern. Measure decision latency, evidence freshness and the cost of wrongly blocked payments together.