Who said what

Where each idea came from, who said exactly what and when, and which popular versions are wrong. Written for saying out loud to someone who may have read the source.

41 ideas, 18 documented wires, 30 misquotes. Every quote here was copied from a page we opened, and a check fails if a word of it changes.

Jump to the loom or to the misquotes.

The loom

A wire appears only where one source names or links another. Press a row to read the words behind each wire. A missing wire doesn’t mean there was no influence. It means nobody wrote it down.

No source in this chain says that two people arrived at something independently. So where dates sit close together and no link exists, this page says the influence is unknown. It doesn’t pick a story.

View
  • Runs on to what it influenced
  • Comes in from what it drew on

Before agents

Older work people now borrow. A wire starts here only where a later page names it.

2025

The words arrive: vibe coding, the lethal trifecta, Ralph.

  6. Geoffrey Huntley

    The Ralph loop, and back pressure for agents

    14 July 2025

    “Anything can be wired in as back pressure to reject invalid code generation.”

    Huntley, Ralph Wiggum as a "software engineer", 14 July 2025

    The written start of the chain. "In its purest form, Ralph is a Bash loop." The post is not the first showing: Horthy records Huntley presenting Ralph at a meetup on June 19th 2025. The word back pressure is older than software.

    Horthy, A Brief History of Ralph, HumanLayer, 6 January 2026

    Taught in the course

    Documented influence

    Then on to: Ryan Lopopolo, OpenAI, Vivek Trivedy, Birgitta Böckeler, Addy Osmani.

    Not on the record

    We found no earlier use of back pressure for agents in this pass, and we did not search exhaustively. No page we opened credits Huntley with coining the agent sense.

    Often misquoted

2026

The practice gets a name, and then an argument about the name.

Date unknown

Misquotes that would embarrass you

The version going round, what the source says, and what to say in its place. Where the record is uncertain, that is the honest answer, and saying it out loud is a point in your favour.

Show

30 of 30

  1. 1People say

    "Agent = Model + Harness" is Hashimoto's, or OpenAI's.

    What is true

    The equation isn't in Hashimoto's post or in OpenAI's. We searched both. The earliest dated appearance we could find is Vivek Trivedy's LangChain post of 10 March 2026, which credits nobody. Böckeler repeats it on 2 April and links that post.

    • “Agent = Model + Harness” Trivedy

    Say instead

    The equation is from Vivek Trivedy at LangChain, as far as I can date it. Hashimoto and OpenAI wrote about harness engineering, and neither wrote that line.

    If they ask where that is from: Trivedy, The Anatomy of an Agent Harness, LangChain blog, 10 March 2026. Add that you can't rule out an earlier use.

    4 sources, and how we read them

    In the course: Harness. On the loom: Vivek Trivedy, Mitchell Hashimoto, Ryan Lopopolo, OpenAI

  2. 2People sayThe record is uncertain

    Hashimoto coined "harness engineering". Or OpenAI did. Or Trivedy did.

    What is true

    Hashimoto's post of 5 February 2026 is the earliest dated use we could find, and the post disclaims any coinage. OpenAI's title came six days later. Trivedy's first dated use came twelve days later, so Osmani's credit to Trivedy doesn't fit the dates. We didn't search exhaustively, and the word harness alone is older.

    • “I don't know if there is a broad industry-accepted term for this yet” Hashimoto
    • “Viv Trivedy coined the term harness engineering” Osmani

    Say instead

    The earliest dated use I could find is Hashimoto, 5 February 2026, and Hashimoto wasn't claiming to have invented it. OpenAI used it six days later. I don't know whether one led to the other.

    If they ask where that is from: Hashimoto, My AI Adoption Journey, step 5. If they bring up Osmani's credit to Trivedy, give the two dates: 5 February and 17 February.

    5 sources, and how we read them

    In the course: Harness engineering. On the loom: Mitchell Hashimoto, Ryan Lopopolo, OpenAI, Vivek Trivedy, Addy Osmani, Justin Young, Anthropic

  3. 3People say

    Huntley wrote "Don't waste your back pressure".

    What is true

    The byline is Moss, 17 January 2026. Huntley's own use of back pressure is in the Ralph post of 14 July 2025. Moss links that post once, on the words "loop agents", and doesn't name Huntley.

    • “Anything can be wired in as back pressure to reject invalid code generation.” Huntley

    Say instead

    Huntley applied back pressure to agents in the Ralph post in July 2025. The essay with that title is by Moss, from January 2026.

    If they ask where that is from: Moss, Don't waste your back pressure, banay.me, 17 January 2026. Huntley, Ralph Wiggum as a software engineer, 14 July 2025, section on phase two.

    2 sources, and how we read them

    In the course: Backpressure. On the loom: Moss, Geoffrey Huntley

  4. 4People sayThe record is uncertain

    Back pressure is a term Huntley invented for agents.

    What is true

    The word began in plumbing, and software had already borrowed it for flow control. Huntley is the earliest use on agents that we found, in a pass that wasn't exhaustive. The agent sense is loose: what gets pushed back is work, and not a flow of data.

    • “a resistance to the desired flow of fluid through pipes” Wikipedia

    Say instead

    It is an old word from pipes and streaming systems. The earliest use I have seen for checks around agents is Huntley's, in July 2025.

    If they ask where that is from: Wikipedia, Back pressure, for the older sense. Huntley's Ralph post for the agent sense.

    2 sources, and how we read them

    In the course: Backpressure. On the loom: Geoffrey Huntley

  5. 5People say

    Everything is a Ralph loop is the Ralph post, and that is where the one-liner comes from.

    What is true

    They are two posts. The Bash one-liner and "deterministically bad" are in the post of 14 July 2025. The January 2026 post has neither, yet it is the one OpenAI and LangChain link. Horthy records Huntley already showing Ralph on June 19th 2025, so the July post isn't the first showing either.

    • “In its purest form, Ralph is a Bash loop.” Huntley
    • “The Ralph Wiggum Technique, created by Geoff Huntley” Horthy

    Say instead

    Ralph is Huntley's. The written description is from July 2025, and Horthy saw it presented a few weeks before that.

    If they ask where that is from: Huntley, Ralph Wiggum as a software engineer, 14 July 2025. Horthy, A Brief History of Ralph, 6 January 2026, for the meetup.

    3 sources, and how we read them

    In the course: Ralph loop. On the loom: Geoffrey Huntley, Geoffrey Huntley

  6. 6People say

    Uncle Bob's Gauntlet is a five-agent product.

    What is true

    No product has that name. It is a figure of speech in one sentence of a post on X, and third-party blogs turned it into a title. The tool is SwarmForge, with packs of two, four and six agents. We found no repo or page of Martin's that names a pipeline Gauntlet.

    • “they’ve had to run the gauntlet of all of my constraints and tests.” Martin on X

    Say instead

    The tool is SwarmForge, which Martin mostly writes. Its six-pack runs specifier, coder, cleaner, architect, hardender and QA. Gauntlet is just a word Martin once used about the checks.

    If they ask where that is from: unclebob/swarm-forge, six-pack branch. For the word, Martin on X, 23 July 2026, and say you read it through a mirror.

    2 sources, and how we read them

    In the course: SwarmForge. On the loom: Robert C. Martin and Justin Martin, Robert C. Martin

  7. 7People sayThe record is uncertain

    Uncle Bob says "I no longer read code", and "they've had to run the gauntlet of all these constraints and tests".

    What is true

    "No longer" is a headline. The research behind this course only had second-hand reports of this. We have since read the post itself, through a mirror of the status, and its wording is a current strategy, about code written by agents. The report we had misquotes its last line ("all these" for "all of my"). "Humans are slow at code" is still only in that report, and we didn't find its original. Martin's spot checks rest on a machine transcript of a talk, so we don't quote them.

    • “My current strategy is to not read any of the code written by my agents.” Martin on X
    • “they've had to run the gauntlet of all these constraints and tests” StartupFortune

    Say instead

    In July 2026 Martin wrote that the current strategy is not to read code written by the agents, and to surround them with constraints instead. Grady Booch disagreed in public, as reported in one article.

    If they ask where that is from: Martin on X, 23 July 2026, a reply, read through a mirror since X blocks fetching. Don't quote "Humans are slow at code" as Martin's unless you have found the post.

    2 sources, and how we read them

    In the course: SwarmForge, Green doesn't mean right. On the loom: Robert C. Martin

  8. 8People sayThe record is uncertain

    Uncle Bob's CRAP threshold is 4 for people and 6 to 8 for agents.

    What is true

    That comes from a machine transcript of a talk, which we couldn't check against the audio. The repo history is better evidence. In SwarmForge the bar went from under 30 (31 March 2026) to 4 (18 April) to 6 (24 May) to 10 on changed files (2 September). ub-mode says under 10 (25 October). Martin's August experiment used 4, and found that it made programs harder to read. A text search of the SwarmForge history found no 8. Savoia and Evans started at 30.

    Say instead

    The published rule now is 10 or below on changed files, with an exception for a single cond or case. It has moved three times this year, so I would quote it with a date, and I would mention that Martin's own experiment found forcing it below 4 made programs harder to read.

    If they ask where that is from: unclebob/swarm-forge, swarmforge/roles/QA.prompt on the six-pack branch. For the caution, unclebob/negative-test-experiment, August 2026.

    7 sources, and how we read them

    In the course: CRAP score. On the loom: Robert C. Martin and Justin Martin, Robert C. Martin, Robert C. Martin, Alberto Savoia and Bob Evans

  9. 9People say

    CRAP is Uncle Bob's metric.

    What is true

    Alberto Savoia and Bob Evans made it at Agitar Labs in 2007, and called the first Java tool crap4j. Martin built later tools for the same formula and uses it as a gate.

    • “my AgitarLabs’ colleague, Bob Evans, and I decided to do some research” Savoia

    Say instead

    CRAP is Savoia and Evans, 2007. Martin uses it as a gate for agents.

    If they ask where that is from: Savoia, Pardon My French, But This Code Is C.R.A.P., Artima, July 2007.

    2 sources, and how we read them

    In the course: CRAP score. On the loom: Alberto Savoia and Bob Evans, Robert C. Martin

  10. 10People say

    CRAP stands for Change Risk Anti-Patterns.

    What is true

    Its authors wrote Change Risk Analysis and Predictions, and the Crap4j site says the same. Martin's crap4clj README writes Change Risk Anti-Pattern, so you will meet both. Only one is the authors'.

    • “CRAP (Change Risk Analysis and Predictions)” Crap4j home page
    • “CRAP (Change Risk Anti-Pattern)” Martin

    Say instead

    Change Risk Analysis and Predictions, as Savoia and Evans wrote it. You will also see Anti-Patterns, including in Martin's READMEs.

    If they ask where that is from: The Crap4j home page, and Savoia on Artima, July 2007.

    3 sources, and how we read them

    In the course: CRAP score. On the loom: Alberto Savoia and Bob Evans

  11. 11People say

    pstack and ub-mode are two people independently landing on the same rules.

    What is true

    Martin says otherwise. ub-mode says it is based on Tan's rules, and its commit history records that note on its first day, 25 October 2026. pstack was first committed on 23 May 2026. The two differ where it counts: Tan's README rejects planning, and Martin's pairs the rules with hard gates.

    • “It is loosely based on portions of @poteto's rules” Martin

    Say instead

    ub-mode is Martin's reduced version of portions of Tan's rules, paired with Martin's own gate tools. That is a documented influence.

    If they ask where that is from: unclebob/ub-mode README, first paragraph. cursor/plugins, pstack folder, for the original.

    2 sources, and how we read them

    In the course: pstack, SwarmForge. On the loom: Lauren Tan, Robert C. Martin

  12. 12People sayThe record is uncertain

    Lauren Tan, whatever-job-title at Cursor, says...

    What is true

    The README gives no title. It gives a history in the author's own words. Don't supply a title you can't source, and don't assume pronouns. Credit the work: Lauren Tan, also known as poteto, wrote pstack.

    • “i've worked with millions of lines of code at Meta, Netflix, and Cursor” Tan

    Say instead

    Lauren Tan, who goes by poteto online, wrote pstack, and it is in the cursor/plugins repo.

    If they ask where that is from: cursor/plugins, pstack/README.md, at version 0.15.36 when we read it. Check the version, because it changes.

    1 source, and how we read it

    In the course: pstack. On the loom: Lauren Tan

  13. 13People say

    pstack's encode-lessons-in-structure principle says the error must name the fix, and that you must prove it fails.

    What is true

    Both of those are in its skill called correct. The principle says to encode the rule rather than write more text. Easy to mix up, and easy to check.

    • “Encode the rule as a lint, metadata flag, runtime check, or script instead of more text.” Tan
    • “Prove each new check fails on a real past mistake.” Tan

    Say instead

    The principle is to encode the lesson in structure. The correct skill adds the two tests: the error names the fix, and the check fails on a real past mistake.

    If they ask where that is from: cursor/plugins, pstack/skills/correct/SKILL.md, and pstack/skills/principle-encode-lessons-in-structure/SKILL.md.

    2 sources, and how we read them

    In the course: Ratchet. On the loom: Lauren Tan

  14. 14People say

    METR proved that AI makes developers slower.

    What is true

    It measured 16 experienced developers on their own open-source repos with early-2025 tools, and they took 19% longer. METR limits the claim itself, and the page now points readers to a February 2026 update.

    • “when developers use AI tools, they take 19% longer than without” METR
    • “We do not claim that our developers or repositories represent a majority or plurality of software development work” METR

    Say instead

    METR's 2025 trial found 16 experienced developers took 19% longer with early-2025 tools while believing they were faster. That is a result about those people, those tools and that year.

    If they ask where that is from: METR, Measuring the impact of early-2025 AI on experienced open-source developer productivity, 10 July 2025. Give the interval: plus 2% to plus 39%.

    2 sources, and how we read them

    In the course: METR's studies, Perception gap

  15. 15People sayThe record is uncertain

    METR's new study shows AI speeds developers up now.

    What is true

    The February 2026 update could not produce a reliable number, because developers would no longer work without AI. Both new intervals include zero. METR believes they are probably faster now, and it calls its own data weak on the size of that.

    • “we believe it is likely a bad proxy for the real productivity impact of AI tools” METR
    • “our data is only very weak evidence for the size of this increase.” METR

    Say instead

    METR's follow-up couldn't get a clean number. Returning developers came out at minus 18%, with an interval from minus 38% to plus 9%. METR thinks speed has probably improved, and says its data can't show by how much.

    If they ask where that is from: METR, We are changing our developer productivity experiment design, 24 February 2026.

    1 source, and how we read it

    In the course: METR's studies

  16. 16People say

    Developer trust in AI went from 31% to 87% in a year.

    What is true

    Stack Overflow's page does make that comparison. The questions differ: 2025 asked how far you trust accuracy, and 2026 asked under what conditions you trust it. The 87% counts everyone who didn't pick distrust or not sure. Only 6.6% trust it for many tasks including important decisions. The 2025 page says 33% trust, which doesn't match the 31% given for last year, and we can't say why.

    Say instead

    Stack Overflow changed the question between years, so the 87% can't be compared with last year's figure. Under the new question, about 7% trust AI on important decisions.

    If they ask where that is from: Stack Overflow Developer Survey 2026, AI section, and the 2025 AI section, for the old question.

    2 sources, and how we read them

    In the course: Trust surveys

  17. 17People say

    DORA found only 3% of developers highly trust AI.

    What is true

    DORA says no such thing. Its figures are 24% for a great deal or a lot, and 30% for a little or none. The 3% is Stack Overflow's 2025 number for highly trusting, in a different survey with a different question.

    Say instead

    DORA 2025 has 24% trusting AI output a great deal or a lot. The 3% is Stack Overflow 2025. They asked different things, so I wouldn't set them side by side.

    If they ask where that is from: DORA, State of AI-assisted Software Development, 2025. Stack Overflow Developer Survey 2025, AI section.

    2 sources, and how we read them

    In the course: Trust surveys, DORA on AI

  18. 18People say

    Three OpenAI engineers wrote a million lines in five months, ten times faster.

    What is true

    Every part of that drops a qualifier. The million is "on the order of", and it includes docs and tooling. The team was three at the start and seven by the end. The tenfold is an estimate with no control. And the post warns against generalising from it.

    • “on the order of a million lines of code across application logic, infrastructure, tooling, documentation, and internal developer utilities” Lopopolo
    • “We estimate that we built this in about 1/10th the time” Lopopolo
    • “should not be assumed to generalize without similar investment” Lopopolo

    Say instead

    OpenAI's team reports about a million lines including docs and tooling, about 1,500 pull requests, three engineers growing to seven, over five months. The speed-up is their own estimate.

    If they ask where that is from: Lopopolo, Harness engineering, OpenAI, 11 February 2026. Mention that it is a team's own report about one repository.

    1 source, and how we read it

    In the course: Software factory. On the loom: Ryan Lopopolo, OpenAI

  19. 19People say

    StrongDM talks about a "satellite of scenarios".

    What is true

    The phrase isn't on StrongDM's page or in Willison's write-up, and a web search turned up no source for it. We don't know where it started. What the page does say is that scenarios sit outside the codebase like a holdout set, and that an agent's shortcut can pass narrow tests.

    • “often stored outside the codebase (similar to a "holdout" set in model training)” McCarthy
    • “return true is a great way to pass narrowly written tests” McCarthy

    Say instead

    StrongDM keeps scenarios outside the codebase, like a holdout set, so an agent can't fit its work to them.

    If they ask where that is from: McCarthy, Software Factories And The Agentic Moment, factory.strongdm.ai, 6 February 2026. Its results are its own report.

    1 source, and how we read it

    In the course: Holdout scenario, Software factory. On the loom: Justin McCarthy, StrongDM

  20. 20People sayThe record is uncertain

    Dex Horthy showed that models get dumb past 40% of the window.

    What is true

    The words are Horthy's, in a dated post. The number isn't: that post says about 75k tokens, and shows no data. The guide says 40 to 60% with no data either. The 40% as a hard line comes from summaries of a talk we couldn't check.

    • “stay in the ~75k token "smart zone" for claude models” Horthy
    • “keeping utilization in the 40%-60% range” HumanLayer

    Say instead

    Dumb zone is Horthy's phrase for a practitioner's rule of thumb. I would clear context early, and I wouldn't quote a percentage as if it were measured.

    If they ask where that is from: Horthy, Context-Efficient Backpressure for Coding Agents, HumanLayer, 9 December 2025.

    2 sources, and how we read them

    In the course: The dumb zone. On the loom: Dex Horthy

  21. 21People say

    Vibe coding means using AI to write code.

    What is true

    Karpathy meant accepting everything without reading it, on throwaway projects, and Willison spells out the difference. Also watch the date. Karpathy's post is from 2 February 2025, and Willison's post says February 6th, so copying from that page gets the date wrong.

    • “I "Accept All" always, I don't read the diffs anymore.” Karpathy on X
    • “Vibe coding is not the same thing as writing code with the help of LLMs!” Willison

    Say instead

    Vibe coding is not reading the code at all. That is fine for a weekend toy. What I do is different, and here are the checks.

    If they ask where that is from: Karpathy on X, 2 February 2025. Willison, Not all AI-assisted programming is vibe coding, 19 March 2025.

    2 sources, and how we read them

    In the course: Vibe coding, Agentic engineering. On the loom: Andrej Karpathy, Simon Willison

  22. 22People sayThe record is uncertain

    Karpathy coined context engineering.

    What is true

    Willison's collection of 27 June 2025 shows Tobi Lütke using it first and Karpathy amplifying it, in Willison's word. Neither claims it. Whoever used it first is unknown to us, and we read these two through Willison and not on X.

    • “Recently amplified by Andrej Karpathy” Willison

    Say instead

    It spread in June 2025 through Lütke and Karpathy. I don't know who used it first.

    If they ask where that is from: Willison, Context engineering, 27 June 2025.

    1 source, and how we read it

    In the course: Context engineering. On the loom: Tobi Lütke and Andrej Karpathy

  23. 23People say

    Kent Beck invented TDD.

    What is true

    Beck's own word is rediscovery. Beck pins the date between SUnit in 1994 and a demonstration in October 1995, and has lost the notebooks that would narrow it.

    • “the timeline of my rediscovery of TDD” Beck

    Say instead

    Kent Beck rediscovered TDD in the mid-1990s, in Beck's own word, and later wrote TDD By Example.

    If they ask where that is from: Beck, Canon TDD, 11 December 2023.

    1 source, and how we read it

    In the course: Test-driven development. On the loom: Kent Beck

  24. 24People sayThe record is uncertain

    Mutation testing is one of Uncle Bob's gates, so it is his idea.

    What is true

    It is decades older than that. DeMillo, Lipton and Sayward published it in 1978, and Wikipedia reports Lipton proposing it as a student in 1971 and Budd writing the first tool in 1980. We could read only the 1978 abstract, on a mirror, and we couldn't open the 1971 paper.

    • “This so-called coupling effect can be used to save work during the testing process.” DeMillo
    • “originally proposed by Richard Lipton as a student in 1971” Wikipedia

    Say instead

    Mutation testing goes back to DeMillo, Lipton and Sayward in 1978. Martin uses it as a gate, and Martin's rule is short: cover the uncovered, kill the survivors.

    If they ask where that is from: DeMillo, Lipton and Sayward, Hints on Test Data Selection, IEEE Computer, April 1978. Say you have read the abstract, if that is the truth.

    2 sources, and how we read them

    In the course: Mutation testing. On the loom: DeMillo, Lipton and Sayward

  25. 25People say

    Brooks proved that software can never get ten times more productive.

    What is true

    The claim is narrower: a single development, within a decade. The reprint says "within a decade" in the opening line, and the IEEE version frames it with "a decade hence". It was also an argued prediction in 1986, and not a proof.

    • “which by itself promises even one order-of-magnitude improvement within a decade” Brooks
    • “as we look to the horizon of a decade hence, we see no silver bullet.” Brooks

    Say instead

    Brooks argued in 1986 that no single development would give a tenfold gain within a decade, because much of the difficulty is in the problem itself.

    If they ask where that is from: Brooks, No Silver Bullet, IFIP 1986, revised in IEEE Computer, April 1987.

    2 sources, and how we read them

    In the course: Coordination cost. On the loom: Fred Brooks

  26. 26People say

    Simon Willison discovered prompt injection.

    What is true

    Willison named it, and says so. The post starts from Riley Goodside's demonstration. It is a fair credit for the name, and it says nothing about who found the attack first.

    • “Riley Goodside, yesterday:” Willison
    • “I propose that the obvious name for this should be prompt injection.” Willison

    Say instead

    Willison named prompt injection in September 2022, after Riley Goodside's example. The lethal trifecta is Willison's too, from June 2025.

    If they ask where that is from: Willison, Prompt injection attacks against GPT-3, 12 September 2022.

    2 sources, and how we read them

    In the course: Prompt injection, Lethal trifecta. On the loom: Simon Willison, Simon Willison

  27. 27People say

    Anthropic found AI users score 17% lower, so AI damages skills long term.

    What is true

    The gap is 17 percentage points, 50% against 67%, which is about a quarter in relative terms. Anthropic's own page says 17% lower, which is how this spread. The quiz came just after one task, with 52 mostly junior engineers. It says nothing about skills fading over months.

    • “scored 17% lower” Anthropic
    • “the AI group averaged 50% on the quiz, compared to 67% in the hand-coding group” Anthropic

    Say instead

    In Anthropic's trial of 52 engineers, the AI group scored 50% against 67% on a quiz straight after the task. That is a learning result, not a long-term one, and Anthropic calls it preliminary.

    If they ask where that is from: Anthropic, How AI assistance impacts the formation of coding skills, January 2026. Mention that the company sells an assistant.

    1 source, and how we read it

    In the course: Skill atrophy

  28. 28People say

    Emil Kowalski says the right words make AI write better code.

    What is true

    Neither page says that. Index says to be precise, and the skill says to go from a fuzzy description to the agreed word. Better work from an agent is this site's claim. Index is credited to Emil and Glenn, so don't credit Kowalski alone.

    • “to go from a fuzzy description of a concept to the precise, agreed-upon word” Index
    • “by Emil and Glenn” Index

    Say instead

    The vocabulary idea comes from Index, by Emil Kowalski and Glenn. The claim that words get you better agent output is this course's, and it tested that on one small model.

    If they ask where that is from: Index, index.how/to/articulate. There is no date or full byline on the page.

    2 sources, and how we read them

    In the course: The right word. On the loom: Emil Kowalski and Glenn, Index, This site

  29. 29People say

    pstack says to run an adversarial review on every change.

    What is true

    pstack never said every change. Its guide says most changes need none of its design steps, and its interrogate skill keeps only findings a run proves. Then on 3 November 2026 its author wrote that they now consider adversarial reviews mostly a waste of tokens, and plan to use the skill less. On 6 November the repo had not changed yet.

    • “Most changes need none of it.” Tan
    • “A finding survives only when a run proves it.” Tan
    • “i now consider adversarial reviews mostly a waste of tokens” Tan on X

    Say instead

    As of November 2026, pstack has an interrogate skill that keeps only findings a run proves, and its author has said they now consider adversarial reviews mostly a waste of tokens. So I would date anything I quote from it.

    If they ask where that is from: pstack 0.15.36, skills/interrogate/SKILL.md, and Lauren Tan on X, 3 November 2026.

    3 sources, and how we read them

    In the course: Adversarial review, Findings must show evidence. On the loom: Lauren Tan, Lauren Tan

  30. 30People say

    Formally verified means it has no bugs.

    What is true

    A proof shows the code meets its spec, under the assumptions it states. Fonseca and colleagues studied three verified distributed systems and found 16 bugs, in the specification, the shim layer and the tools, and none in the verified protocols.

    • “we found a total of 16 bugs” Fonseca
    • “none of these bugs were found in the distributed protocols of verified systems” Fonseca

    Say instead

    Verified means it meets its spec under stated assumptions. The bugs that remain live in the spec and in everything around the proof.

    If they ask where that is from: Fonseca, Zhang, Singh and Ram, An empirical study on the correctness of formally verified distributed systems, EuroSys 2017.

    1 source, and how we read it

    In the course: Formal verification. On the loom: Pedro Fonseca and colleagues

What we couldn’t open

The words each quote was checked against are in docs/research/excerpts/, one file per source, and bun scripts/check/lineage.ts runs with the other checks behind the receipts. Pages change. These were read on 5 November 2026.

Before the agent

  1. 1The machine
  2. 2The network
  3. 3Version control
  4. 4Reading and debugging

The agent

  1. 5The model
  2. 6The agent
  3. 7Context
  4. 8The brief

The harness

  1. 9Backpressure
  2. 10Tests
  3. 11Design
  4. 12Review

Everything it can get wrong

  1. 13Security
  2. 14Data and failure
  3. 15Performance and cost
  4. 16Shipping and running

The operator

  1. 17Many agents
  2. 18The proof

The workshop

Name a word, or describe what you mean in everyday language.