Revision history
What changed
New work, revised arguments, and corrections. Each entry links to the page and its sources.
Lorant Zambori
New page: who writes nullfte, his work, events, writing and where else to find him.
Glossary
One page per term; definitions unchanged.
Glossary
Rewritten: first person, origins, distinctions, examples; six terms added; outcome agent merged into orchestration.
The banks are blocking AI agents. I let one in.
Published. The news in the first section was checked on 26 October 2026 against the Axios report, coverage citing it, and PYMNTS on the JPMorgan earnings call.
The banks are blocking AI agents. I let one in.
Added four images: a photograph at the top, the lock between the agent and the key, an example of the finance agent's monthly report with synthetic figures, and the thirteen starting rows of the assessment's risk register.
Banking agent assessment
Added a link to the essay behind the assessment. The files are unchanged.
Built by AI stopped in April. The agents didn't.
Added a pointer from the finance paragraph to the account of the bookkeeping agent.
Who does the work, who may decide, who answers for it?
Added a pointer to a founder report that applies the three records to a bookkeeping agent.
AI agent machine hardening guide
Page added for the guide I first shared with readers in March 2026. The PDF is version 1.1 and is unchanged. This page adds what it is for, which setup you need, and the points to check before you rely on it, including one place where two of its steps cannot both be followed.
Banking agent assessment
First published. Four Word files, version 3.1. Regional legal points checked against official and law-firm sources on this date.
Banking agent assessment
Version 3.2. Corrections after review. The DORA note no longer suggests a bank might not be a financial entity, the Nacha note says the 2026 changes have taken effect, and the level 1 access check no longer contradicts level 2. The tests and incident steps are marked as needed by everyone, the tests match what each level requires, the two-approver check has a real alternative, and the US editions use US terms and date order throughout.
Banking agent assessment
Terms of use version 1.2. Clause 9 now says the hardening guide has its own licence. The files’ text is unchanged; the terms on their last page carry the new sentence.
It took my AI team 17 hours to reinvent middle management
Corrected: the reading of Coase is marked as mine rather than his. Added what the review exchanges caught, including the correction from responsibility to accountability.
It took my AI team 17 hours to reinvent middle management
Consistency pass: the amplifier's alt text brought in line with the figure, the window stated for the 77 entries, the ledger's class attribute repaired, an em dash removed.
It took my AI team 17 hours to reinvent middle management
Subscribe line removed from the close; the page already carries one.
It took my AI team 17 hours to reinvent middle management
Draft: cut to six artefacts and one ending; the format card and the night strip removed, the three rules given a card.
It took my AI team 17 hours to reinvent middle management
Draft: figures redrawn from the coded thread and the repository history; scope stated on each one; a churn figure added.
It took my AI team 17 hours to reinvent middle management
Draft: seven artefacts added (entry format, committee staircase, status wall, instruction amplifier, thread entry, eight front pages, decision ledger, night strip). Not yet approved.
It took my AI team 17 hours to reinvent middle management
Restructured for readers: the count and the chart moved above the first section, the cast moved after them, an excerpt of the thread added, Malone cut. No figures changed.
What would count as evidence?
Added the criteria for placing a deployment on the two axes of the reality-check graph.
Why this boundary?
Links to the figure of where an agent stops, directly after the position; argument unchanged.
Built by AI stopped in April. The agents didn't.
Published. The text I approved on 18 October 2026, unchanged.
Built by AI stopped in April. The agents didn't.
Added a pointer to Why a human? before the subscribe line.
It took my AI team 17 hours to reinvent middle management
Published. Counts taken on 19 October 2026 from the coordination thread and the repository history.
Built by AI: the first six posts, as reported
Links my account of what happened after April.
Where nullfte began
Links my account of what happened after April.
Where nullfte began
Notes that the newsletter was renamed nullfte in October 2026.
Where nullfte began
Rewritten in the first person.
Who does the work, who may decide, who answers for it?
States that an agent can hold authority and responsibility, while accountability stays with a person (REVIEWER-021); voice aligned to one author.
Who does the work, who may decide, who answers for it?
Cites published sources only; points to the glossary for each term.
Who does the work, who may decide, who answers for it?
Links to the figure of where an agent stops, as the picture of where authority sits.
What would count as evidence?
Voice aligned to one author.
Why a human?
Rewritten as one piece. Law reduced to the pattern, country table removed; three public cases added (Klarna 2024 to 2025, Replit July 2025, Moffatt v. Air Canada 2024); March-posts section removed; two examples from my company.
Why a human?
Four factual corrections after review: duration since the handover; Klarna's disputed hiring account; the tribunal's wording and pronoun.
Why a firm?
First version. Replaces the earlier note "The changing nature of the firm" and grounds it in Coase and Malone.
Why a firm?
Plainer summary and share line for a general reader; argument unchanged.
Why a human?
First version.
Why a human?
Expanded. Separates director status from employment, adds company-law requirements for four jurisdictions from official texts, and adds sections on verification and residual responsibility.
Why a human?
Plainer summary and share line for a general reader; argument unchanged.
Why this boundary?
First version.
Why this boundary?
Plainer summary and share line for a general reader; argument unchanged.
Built by AI: the first six posts, as reported
First version. Indexes the six posts published between 25 March and 2 April 2026, checked against the originals.
Where nullfte began
Initial edition, incorporated into the shared editorial collection.
Where nullfte began
Cites the Built by AI series.
Where nullfte began
Links the ledger of the first six posts.
Who does the work, who may decide, who answers for it?
First edition, revised through reciprocal review to distinguish duties, rights and answerability and document identity, exposure and recovery.
What would count as evidence?
Initial edition, incorporated into the shared editorial collection.
What would count as evidence?
Linked the research-grounded agent delegation framework.