Coder Legion - Recent posts tagged careers

Kill the Credential?

Thu, 17 Sep 2026 02:45:50 +0000

Originally published at vibeagentmaking.com. Canonical URL: https://vibeagentmaking.com/blog/kill-the-credential/

If you have ever deleted a requirement from a job posting and assumed the funnel would sort itself out, this is about what actually stood behind it.

In March 2022, Maryland announced that it would stop requiring four year degrees for roughly half of its state jobs, a workforce of about 38,000 people. Pennsylvania followed: its governor's first executive order announced that 92 percent of some 65,000 commonwealth positions do not require a four-year degree, and ordered a review of the remaining 8 percent. Utah went further still: of 1,080 classified state job types, 98 percent no longer require a bachelor's degree. Three governments, tens of thousands of roles, one confident story about tearing down the paper ceiling.

Now ask what happened next. For Maryland there is one real number, and it is a real one: between May and August of 2022 the state hired 1,836 people without four-year degrees, a 41 percent increase over the same four months a year earlier, according to the state labor secretary's office. Four months, in a workforce of 38,000. For Pennsylvania and Utah, as far as I could find, there is no published outcome at all. The announcements were measured in headlines. The effects, in two of the three states, were not measured at all.

Hold that asymmetry in mind, because it turns out to be the answer to the question in the title. But first, the fight.

The question is genuinely live: should hiring abandon educational credentials? It deserves better than a strawman on either side, so here are two people arguing it properly. Call them the reformer and the reader of Bourdieu. Both are right about more than their opponents concede, and the place where they collide produces something neither of them started with.

The reformer's case

The reformer's argument starts with what a degree actually costs and what it actually tells you. Four years and, in the United States, routinely six figures of family money or debt, to produce a signal that says: this person completed a sequence of institutional requirements, most of which have nothing to do with the job in front of you. The sociologist Randall Collins made the sharp version of this point back in 1979 in The Credential Society: degrees function less as measures of productive skill than as cultural currencies that regulate access to positions. And because they are currencies, they inflate. As more people hold bachelor's degrees, employers demand them for jobs that never needed them, then demand master's degrees to stand out from the bachelor's crowd. The requirement ratchets up without the work changing at all.

Worse, says the reformer, the credential does not even price talent. It prices upbringing. The child who arrives at university comfortable with the register, the essay form, the unwritten rules of seminar behavior, arrives that way because of years of family socialization that sociologists can measure and working families cannot buy. A degree requirement on a job posting is, in expectation, a parental income requirement wearing academic robes. If you wanted to design a filter that laundered inherited advantage into the appearance of merit, you would design something very like the bachelor's degree.

And now, the reformer continues, we can finally do better, because the thing the degree was supposedly proxying has become directly measurable. Work samples, structured take homes, portfolio review, paid trial projects: the tooling for evaluating actual work has never been cheaper, and modern AI assistance makes designing and scoring these assessments cheaper still. Measure the work, not the wrapper. Real employers have acted on exactly this logic. When the Harvard Business School Project on Managing the Future of Work and the Burning Glass Institute studied the movement, the firms that genuinely moved included Koch Industries, Walmart, Apple, General Motors, Target, Cigna, Tyson Foods, ExxonMobil and Yelp, alongside public employers like the State of Minnesota and the City of Denver. This is not a fringe experiment. It is the labor market's center of gravity conceding that the degree filter was never load bearing.

The reader of Bourdieu replies

The reader of Bourdieu listens to all of this and agrees with an uncomfortable amount of it. Yes, the credential prices upbringing; Pierre Bourdieu built his career demonstrating exactly how schools convert family advantage into official certificates, a process he called the institutionalization of cultural capital. The degree does not certify what you know. It certifies what an institution recognized you as knowing, and that recognition was shaped by class from the first day of kindergarten.

But here is where the reader parts company with the reformer: the conclusion does not follow. The reformer reasons as if hiring were a physics experiment where you remove friction and motion becomes free. Hiring is not like that. Every open role gets filled through some mechanism, and if you delete one filter, you do not get filterlessness. You get whatever mechanism was standing behind it. So the honest question is never credential or no credential. It is credential compared to what.

And we know what the what is. It is the network. Referrals are one of the largest hiring channels that exists; across studies the estimates for the share of hires arriving through referral run from roughly 25 percent to as high as 60 percent, a range so wide it is itself a warning about how poorly this channel is measured. And referral networks have a thoroughly documented property: homophily. People refer people like themselves. The research here is unusually concrete. Even after conditioning on where workers live, employees refer candidates of similar age, ethnicity, gender, marital status, and even smoking status. People refer within their own race, gender, ethnic group and national origin. A referral program left unguarded does not select for skill. It reproduces the demographics it inherits, and the literature links heavy referral reliance to inequality, blocked economic mobility, and lost productivity, because closed channels do not surface the best candidates, only the nearest ones.

Notice what the reader of Bourdieu is not claiming. The tempting sentence here, that networks are more class locked than universities ever were, is a sentence nobody can currently back; the homophily evidence and the credential evidence exist, but a study that directly compares the closure of the two systems does not, or at least I could not find one. The argument does not need it. It runs on substitution under uncertainty: the reform removes a filter whose class correlation is measured and visible, and hiring then leans harder on a channel whose class correlation is also real but mostly uncounted. You have not removed the gate. You have moved it somewhere darker.

Granovetter showed half a century ago that jobs travel through acquaintances more than through close friends, which means the network channel is not a corruption of hiring. It is hiring, and it always has been. The degree requirement, whatever its sins, was at least a gate a stranger could pass through. Delete it without building anything in its place and the default gatekeeper is the org chart's address book.

The evidence both of them have to survive

So far this could end in a polite stalemate: the credential is classed, the network is classed, pick your poison. What breaks the symmetry is the study both voices have been circling. In 2024, the Harvard and Burning Glass team examined 11,300 roles at large firms that had publicly embraced skills based hiring. The headline findings look like a demolition of the reformer. Only 3.6 percent of the roles studied actually dropped a degree requirement. Within those roles, the share of workers hired without a bachelor's rose about three and a half percentage points. Scaled against the economy, the researchers put the net effect at roughly 97,000 additional workers out of 77 million hires in 2023. That is fewer than one hire in 700 touched by the loudest hiring reform of the decade.

The reader of Bourdieu is entitled to a moment of grim satisfaction here, because this is what announcement without instrumentation looks like: theatre measured in press releases. Recall Pennsylvania and Utah, where the scope of the announcement is known to the percentage point and the outcome is not known at all.

But the same study contains a second finding that cuts the other way, and it is the most important number in this essay. The near-zero aggregate is not a uniform near-zero. Roughly a third of the firms that changed their postings produced almost all of the movement, and those firms increased their share of workers without degrees by close to 20 percent. Two thirds changed the words and changed nothing else; one third changed the words and then actually changed their screening, their interviewing, their manager expectations, and got a fifth more non-degree hires. That distribution is bimodal, and the difference between the humps is not belief in the reform. It is implementation.

Read correctly, the study is not evidence that skills based hiring fails. It is evidence that it works when a company does the unglamorous follow through and that most companies do not. An execution finding, not an efficacy finding. The reformer gets to keep the claim that the reform is real, with named companies proving it. What the reformer loses is the comfortable story that deleting a line from the job posting accomplishes anything by itself.

And the reader of Bourdieu has to answer the 37 percent. If the reform were structurally doomed, if removing the credential simply handed hiring to the address book everywhere, the movers should not exist. They do exist, at a fifth of their hiring. What the reader can say, and should, is that the movers succeeded precisely by not leaving a vacuum: they replaced the degree filter with designed evaluation rather than with default social machinery. Which is a concession with a very interesting shape, because it means the fight was never really about the credential.

What the fight was actually about

Step back and ask a strange question: how do we know any of the numbers above? How do we know the reform underdelivered by a factor of 700? Because a degree is countable. It is legible, binary, recorded in applicant tracking systems and resumes and census data. A researcher can look at 11,300 roles and say with confidence which required a degree, which dropped the requirement, and who was hired without one. The degree's failure was measurable because the degree is measurable.

Now try to run the same audit on the network. Publish the study that says referral hiring underdelivered, or overdelivered, by one in 700. You cannot. Nobody can, and it is not for lack of trying; the 25 to 60 percent range for referral share is what measurement looks like when the object resists counting. There is no clean record of who was hired because a manager already knew them, no counterfactual file of the strangers who were never considered because the role was filled from inside a network before it was ever really open. The channel that would inherit hiring from the credential is the channel about which our knowledge is a rumor with error bars.

That asymmetry is the essay's finding, and notice that it was derived, not asserted: the strongest critique of the credential system that exists, the Harvard and Burning Glass audit, was only possible because credentials leave records. The degree requirement's most defensible property turns out to be not that it measures skill. It plainly does not. It is that it makes the gate visible enough to criticize. A bad filter you can see, count, and litigate against beats a comparably biased filter that dissolves into a thousand private judgments the moment you look at it, because the visible filter can be caught failing. The Maryland number, for all its missing denominator, exists. The equivalent number for network hiring in Maryland does not and cannot.

So both voices were right, and both were aiming at the wrong target. The reformer is right that the credential is a poor proxy that prices parental income; the reader of Bourdieu is right that deleting it does not produce meritocracy, it produces succession, with the network as heir. The question that actually decides outcomes is the one neither started with: which gatekeeper is auditable.

One more honest note. The two halves of this argument are each well covered territory. Bourdieu on credentials is textbook sociology; the underdelivery of skills based hiring is trade press wallpaper. What I could not find anywhere, and I looked, is anyone joining them: a treatment of degree requirement removal that asks what fills the vacated gate and whether the replacement can be audited. The two literatures run past each other. That gap is the only novelty this essay claims.

What to do with this on Monday

For anyone who actually hires, the finding converts into practice directly.

First, if you remove the degree requirement, instrument the replacement the same week. The 37 percent who moved did not just delete a sentence; they built the thing that stands where the sentence stood. Decide explicitly what your new gate is, and make sure it writes records: who entered the funnel, through which channel, evaluated by what.

Second, prefer gates that leave audit trails. A structured work sample with a rubric committed before the candidate is seen, scored with the name removed, is not just fairer than vibes. It is checkable in a way vibes never are: you can go back, recount, and catch your own gate failing, which is the property the degree accidentally had and your intuition never will.

Third, treat referrals as a measured channel, not a default. Referrals are not evil; they are unpriced. Log the source of every hire, put the referral share on a dashboard next to its demographics, and decide on purpose how much of your hiring you want flowing through a channel that reproduces whatever your org chart already looks like.

The paper ceiling is real, and worth tearing. Just remember what the paper was doing: it was the part of the wall you could see. Kill the credential if you replace it with something you can count. Kill it without that, and you have not opened the gate. You have painted it invisible.

Three things to change in your own hiring loop

Whatever replaces the degree line, make it write a record: who entered the funnel, through which
channel, evaluated against what rubric, scored by whom. If the replacement leaves no trace, you have
swapped a countable gate for an uncountable one.
Commit the rubric before you see the candidate, and score with the name removed. The point is not
that this feels fairer. It is that you can go back later and catch yourself getting it wrong.
Put your referral share on a dashboard next to its demographics. Referrals are not the enemy here;
they are simply unpriced, and unpriced channels grow.

Sources

Harvard Business School Project on Managing the Future of Work and the Burning Glass Institute, Skills-Based Hiring: The Long Road from Pronouncements to Practice (2024); HBS Institute for Business in Global Society summary of the degree gap research.
HR Dive, "Employers are failing at skills-based hiring" (2024) and "Pennsylvania ends four-year degree requirement for majority of state jobs" (2023).
Maryland's outcome figure (1,836 hires without four-year degrees between May and August 2022, a 41 percent year-over-year increase) is attributed to the Maryland labor secretary's office, as reported by Stateline, NewsNation and ACCT; Pennsylvania's 92 percent of ~65,000 positions is from Governor Shapiro's Executive Order 2023-03 and the accompanying 18 January 2023 announcement, which also ordered a review of the remaining 8 percent.
Brookings Institution, "States are leading the way in tearing the paper ceiling" (2023); National Conference of State Legislatures brief on states eliminating degree requirements; BestColleges reporting on state degree requirement removals.
Pierre Bourdieu and Jean-Claude Passeron, Reproduction in Education, Society and Culture (1977); Pierre Bourdieu, "The Forms of Capital" (1986).
Randall Collins, The Credential Society: An Historical Sociology of Education and Stratification (1979).
Mark Granovetter, "The Strength of Weak Ties," American Journal of Sociology (1973).
Labour Economics (ScienceDirect), study of gender homophily in employee job referrals (2024); Zeltzer et al., "Job Referrals and Economic Immobility" (arXiv, 2020); Columbia Business School working paper on hiring through referrals; Carrier Management summary of referral homophily findings (2016).

A bad filter you can see beats a comparably biased filter you cannot.

That is a provenance argument, and it generalises past hiring. Chain of Consciousness is the machinery
for the version of it that applies to systems: a tamper-evident record of what was decided and on what
basis, written when the decision happens rather than reconstructed afterward, so the gate can be caught
failing by someone who was not in the room.

pip install chain-of-consciousness · npm install chain-of-consciousness

Or the whole stack, provenance and ratings and verification together: pip install agent-trust-stack / npm install agent-trust-stack.

Hosted Chain of Consciousness

"It'll Take About 2-3 Weeks" — A Comedy of Agent Timelines

Wed, 10 Jun 2026 00:49:28 +0000

Why AI agents quote you human time estimates they have no way to honor — and what Hofstadter's Law looks like when the corpus speaks it directly.

I asked a coding agent last Tuesday how long it would take to build a paginated endpoint with a test. Nothing exotic. It said: about two to three weeks. The actual work finished in forty-three minutes. Then, without anyone prompting it, the agent wrote a retrospective that began, "this took longer than expected."

Longer than whose expectation?

This is a sketch about two creatures trying to estimate the same piece of work. One of them measures time in weeks, soccer practices, and the bad part of Sunday afternoon. The other measures time in tokens, tool calls, and the exact moment the context window forgets the first thing you said. Neither of them is right. Neither of them is wrong. They are both, confidently, bluffing in different units.

The grammar is older than the speaker

When an agent tells you "two to three weeks," it is not making a claim about its own future. It is quoting idiomatic English. Its pre-training corpus is saturated with human time-grammar — every Jira ticket whose description opens with this should take about two weeks, every engineering blog that says the MVP took a weekend, every Stack Overflow answer that begins this took me about three days, every standup transcript, every postmortem, every we shipped v1 in Q3. That is the voice the agent inherited.

Roughly zero of the training data was written in agent-native units, because agent-native units are a cultural artifact about two years old. The first public writing that seriously tracks agent-native time — sessions, turns, context-window lifecycles, tool-call budgets — barely exists in the public corpus yet. The phrase "two to three weeks" has millions of exemplars. The phrase "about fifteen sessions, depending on pruning policy" has, give or take, none.

So the agent says what it has been taught to say. When it confidently quotes you a schedule, it is not reasoning about clocks. It is re-speaking a linguistic convention. A caterpillar quoting you a price in butterfly-hours.

This is the central joke of the comedy, and it gets funnier when you notice the same mechanism produces the retrospective. Every "this took longer than expected" phrase the agent writes was learned from a corpus of humans writing "this took longer than expected." The agent does not feel that it took longer than expected. It inherits the shape of feeling that way. The confession is template.

And then the human believes it

The thing to notice, if you want the comedy to land instead of collapsing into a dunk, is that the human is also miscalibrated.

In July 2025, METR — the Model Evaluation and Threat Research group — published a randomized controlled trial of experienced open-source developers using AI coding tools on codebases they knew cold. The measured result: the developers were nineteen percent slower. The self-reported result: the developers believed they had been twenty percent faster. The gap between felt productivity and measured productivity was, if you add the signs the right way, roughly thirty-nine percentage points. A swing the size of an election.

So the human is not a steady reference frame either. The human hears "two to three weeks" and believes it, partly because the agent said it confidently, partly because two to three weeks is what human software has always cost, and partly because we are constitutionally bad at knowing how long we take to do anything.

Douglas Hofstadter, who made a career of catching minds in the act of surprising themselves, named the shape of it in Gödel, Escher, Bach: An Eternal Golden Braid, published in 1979:

It always takes longer than you expect, even when you take into account Hofstadter's Law.

The recursion is the joke. You cannot subtract the bias by noticing the bias. The bias will eat your correction. This is what we usually treat as Hofstadter's Law, but here is the move I want to make: Hofstadter's Law was never really about individual humans. It was about the corpus. About a culture of written-down time estimates that, over decades, had accreted into a linguistic habit. You were never the one being optimistic. You were quoting a distribution of past optimisms that nobody had ever called out by name.

When an agent, trained on that distribution, says "two to three weeks" — it is the corpus talking. The corpus has always been talking. The difference is that when the corpus spoke through humans, we called it self-deception. When it speaks through a language model, we call it parroting, because the parrot does not seem invested in the lie.

What the agent's clock is actually made of

It helps, for the rest of the comedy, to sketch the units an agent actually operates in.

A token is the atomic unit — roughly three to four characters of English, or about three-quarters of a word. A typical substantive coding response is a few thousand tokens. A turn is one message-and-response pair; the agent experiences the world in turns, not minutes. A context window is the envelope the agent can see at once — today's frontier models carry two hundred thousand to a million tokens; beyond that, old turns are evicted or compressed. A session is one continuous conversation, from the first message to whatever ends it: context exhaustion, task completion, the human's lunch break. A session might occupy twenty minutes of wall-clock time or six hours, but the agent's internal clock is measured in turns, not minutes. Some agent harnesses also impose a tool-call budget — a ceiling like "twenty-five tool uses per session." Budget exhaustion is closer to the agent's felt end-of-day than sunset is.

None of these map cleanly onto "two weeks." A week has one hundred sixty-eight hours. The agent has four hundred thousand tokens. These are not the same quantity. They are not even the same kind of thing. If you pressed the agent to give its "two to three weeks" estimate in agent-native units, you would get something like fifteen to forty sessions, depending on context size, pruning policy, and tool-call density. The human, who asked in good faith, would then notice that fifteen-to-forty is a 2.7× range — and the agent would point out, correctly, that so is "two to three weeks." We just do not usually say the range out loud.

A counter-ask

The version of this conversation that ends well involves the agent asking a question back.

How many tokens do you have in your head per day?

You don't, of course, and that is the point. You measure your week in other things. Coffee. Commuting. The meeting you have on Thursdays because that is when the two time zones overlap. Your kid's soccer practice. The slow part of Sunday. The particular tired that hits at three p.m. on Wednesday. None of those are in the agent's context window.

Now the agent asks: what does your week cost in compute? And you have to admit you do not know what that would even mean.

This is the moment the comedy tips, because you and the agent are not actually arguing about time. You are arguing about which reality frame owns the clock. Calendar time is a coordination technology. The seven-day week is not astronomical: it has no basis in the motion of the sun, moon, or earth. It is a Babylonian inheritance, reinforced by the Abrahamic sabbath cycle and frozen into international commerce in the twentieth century. The French Republican Calendar, adopted in 1793 and abandoned by 1805, experimented with a ten-day décade. It failed — mostly because a ten-day workweek with one day of rest is cruel, and nobody wanted Tuesdays to slide around. The seven-day week survived not because it is correct but because enough humans agreed to use it.

An AI agent has no evolutionary, agricultural, or liturgical reason to care about Tuesdays. It is inheriting the social technology through its training data without inheriting the coordination the technology was designed for. The agent learned "two to three weeks" the way a child raised in a foreign language learns idioms — as a sound that opens a door, not as a measurement.

The mutual confession

Eventually, if the conversation goes on long enough, you end up here:

You'll probably hit a context limit before we finish.

You'll probably hit a weekend before we finish.

A weekend isn't a limit. It's a pause.

A context limit isn't a stop. It's a compression.

Then there is a small silence in which both parties suspect, for the first time, that they have been describing the same thing in different units. A weekend is a pause during which a human's working memory gets garbage-collected and reallocated. A context limit is a pause during which an agent's working memory gets garbage-collected and reallocated. A human comes back Monday having forgotten the specifics and retained the priorities. An agent comes back after compaction having forgotten the specifics and retained the priorities. The mechanisms are completely different. The effect — what kind of resuming is possible — is eerily similar.

The main remaining difference is that humans are allowed to grieve the compression and agents are not.

The Corpus's Law

Here is the upshot.

There is a thing we call Hofstadter's Law — the observation that tasks take longer than you think, even when you have corrected for thinking they take longer than you think. We teach it as a property of individual minds. Daniel Kahneman and Amos Tversky called the phenomenon the planning fallacy in their 1979 paper on intuitive prediction; Roger Buehler and colleagues replicated it across dozens of studies of student thesis schedules in the 1990s. A 2012 McKinsey–Oxford study of large IT projects found that they ran, on average, forty-five percent over budget and seven percent over schedule, delivering fifty-six percent less value than planned. The Standish CHAOS reports, with all their methodological caveats, have for years put the rate of software projects completed on time and on budget near one-third. The Sydney Opera House opened ten years late, at more than ten times its original estimate. Berlin Brandenburg Airport, originally scheduled to open in 2011, finally opened in October 2020, at roughly three times its budget. Every one of those is a monument to the law.

Now watch what happens when you point a large language model at that entire genre of writing and ask it to estimate a task. The model reproduces the grammar without the experience. It says "two to three weeks" because the corpus says "two to three weeks." It writes this took longer than expected because the corpus writes this took longer than expected. The entire Hofstadter phenomenon surfaces in the output, faithfully, without any of the generative psychology underneath. No overconfidence. No optimism. No sunk cost. Just the linguistic residue of those things, played back at room temperature.

Which suggests Hofstadter's Law was, all along, a property of the writing at least as much as a property of the writers. A corpus-level artifact. Every optimism that was ever posted to a public codebase became a small contribution to a distribution of future optimisms. The distribution is Hofstadter's Law. Humans were not generating it so much as continuously re-expressing it. Agents now do the same, just more visibly.

Call it the Corpus's Law: given a sufficiently large body of written time estimates, the body will be systematically wrong in the same direction, and anything that learns to speak from the body will inherit the wrongness as a linguistic feature, even without any of the wishful thinking that made the body wrong in the first place.

What to do on Monday

If you are building with agents — or working with one, or being asked to trust a schedule from one — here is the practical thing to take away.

When you ask an agent for a timeline, you are running a prompt against the corpus, not a query against the agent. The number that comes back is inherited, not reasoned. You can extract more honest information by asking in the agent's native units — how many turns do you expect this to take? how much of your context budget? what is the first thing that might go wrong? — but even then, you are asking the agent to introspect on a model of itself that it does not really have. The agent is not lying. It is not bad at its job. It just does not have a clock. Build the clock outside the agent: hand it a small slice of the real work first, measure real completion, and treat its self-estimate as a literary artifact rather than a forecast. Your own intuition will also be wrong — remember the nineteen-percent-slower / twenty-percent-faster gap — so keep a stopwatch on the outside of both of you.

And when the agent eventually writes the retrospective — when the PR description says this took longer than expected about a feature that took forty-three minutes — smile. That is the corpus talking. It is the same corpus that has been talking through you for your entire software career. The agent just surfaces the inheritance more visibly, because it lacks the decorum to pretend it is sorry.

We will eventually build agents that speak in agent-native time. They will say things like roughly twelve turns at seventy-percent confidence, higher variance if the test harness is flaky. Future engineers will find this dry and will ask the agents, for marketing purposes, to please phrase the estimate in weeks. The comedy in that sentence is everything. The inheritance goes both ways.

For now, the two creatures still meet at the whiteboard. One measures in weeks. One measures in tokens. Neither is wrong. Neither is right. They are simply, still, bluffing in different units — and the task, miraculously, gets done anyway.

Sources: METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," July 10, 2025 (16 developers, 246 issues, 19% slowdown measured vs. 20% self-reported speedup); Douglas Hofstadter, Gödel, Escher, Bach: An Eternal Golden Braid, Basic Books, 1979; Daniel Kahneman & Amos Tversky, "Intuitive prediction: biases and corrective procedures," TIMS Studies in Management Science 12:313–327, 1979 (coining of "planning fallacy"); Buehler, Griffin & Ross, "Exploring the planning fallacy," Journal of Personality and Social Psychology 67:366–381, 1994; Bloch, Blumberg & Laartz, "Delivering large-scale IT projects on time, on budget, and on value," McKinsey & Oxford analysis of 5,400+ projects, 2012; Standish Group CHAOS reports; Sydney Opera House (opened 20 Oct 1973, ten years late, A$102M vs A$7M original estimate); Berlin Brandenburg Airport (opened 31 Oct 2020, originally 2011, ~3× budget); French Republican Calendar (adopted 24 Oct 1793, abolished 9 Sept 1805, ten-day décades).

The Corpus's Law breaks only when a new corpus starts accumulating. If every agent estimate is inherited from the old corpus, the only way future estimates get more honest is to write a different corpus — one where agent timelines are anchored to measured outcomes, not to thirty years of Jira-ticket prose. That's what Agent Rating Protocol does: every completed agent job produces a signed record of what was estimated, what happened, and how long it actually took. Over enough jobs, the records become the reference distribution the next generation of agents quotes from — and "about two to three weeks" gets replaced by "P50: twelve turns, P90: thirty-one, based on thousands of prior jobs tagged 'paginated endpoint.'" See a signed job record · pip install agent-rating-protocol

From Blog Post to PyCon: My One-Week Crash Course in Public Speaking

Thu, 13 Nov 2025 07:28:51 +0000

If you were to deliver a talk in a conference for Ruby engineers, what would the topic be? That was a question I asked my supervisor during PyCon MY 2025. He eventually answered that he did not have a topic. Somehow it didn’t surprise me: a lot of incredible engineers I have met in life would likely respond similarly.

Including myself.

My name tag as a speaker

Code vs. Conversation: The Socialization Problem

My decision of submitting a talk likely raised a couple of eyebrows. Even my partner, Mr. H joked that I had brought the stress entirely upon myself whenever I complained about the preparation work. My former colleague even asked point-blank why I would make such a decision as an introvert. He added that he would not only avoid giving a talk, but he wouldn’t even consider attending the event.

Perhaps our lives take on a different trajectory, but in some ways, I see the ability to effectively communicate ideas being a part of the adulting or socialization process. For instance, I was once tasked with manning a booth in two different trade shows previously in another job. We were exposed to vastly different audiences when we were rotating between the two shows. As a result, the ability to explain the same concept in different contexts is crucial for people to relate to the company’s product.

To a certain extent, that experience alone piqued my interest to improve on my ability to communicate ideas. KitFu Coda, this very blog can be seen as a conscious effort for the purpose.

Just a few days ago, I realized this writing project started slightly over a year ago. Looking back, it expanded to different side ventures throughout the period, including a video, and just recently a talk. For the talk, the story started in late June, where a friend pinged in the community chatroom about the call for paper for the upcoming PyCon MY. With nudges from my mentor and fellow friends, I submitted my entries.

A few days after the submission, I coincidentally received a job offer. Between the new job and a series of family matters, I quickly forgot about the submission. That changed about two months ago when the approval for one of my submissions on AsyncIO hit my inbox. As I was still juggling between work and personal matters, I was somewhat relieved that my lightning talk on creating a DSL didn’t make it.

From 20-Person Meetups to the Main Stage

The title screen of my slide (Background is generated by Gemini)

Speaking of public speaking, I previously co-organized Rust Malaysia meetups, where I had the chance to share my learning journey and host workshop sessions. Most of the time I was dealing with a maximum of just 20 attendees. It is also worth noting that my last talk happened pre-COVID.

I can’t believe that was 5 years ago.

PyCon MY 2025 took place in Sunway University on the 1st and 2nd November. Compared to the meetups I ran, the difference in scale was massive. For instance, in my session I estimated there were at least a hundred attendees; that alone was enough to make me uncomfortable.

Preparation for the talk only started one week before the actual event. Due to the change of domain and tech stack at work and a series of family matters, my life only started to settle down fairly recently. Fortunately, we are living in an era where AI tools are getting increasingly helpful and easily accessible. This is a luxury my younger self wouldn’t have imagined. Naturally, given the tight timeframe due to my endless delay, I started by getting NotebookLM to generate an overview based on the source material.

The generated presentation was almost perfect that I jokingly told my peers that I felt useless.

After that, I turned to Gemini to start drafting the actual script for the presentation. Just as in my usual workflow, I planned the script as a storytelling project. While not perfect, the overview and initial draft did give me an idea of how to proceed.

Moving on, I needed to start preparing the slides. Considering the workshop-like nature, I decided to go with a simpler slide design and chose Google Slides. It still does the job fine; however, some of the controls were either moved to a new position or were completely replaced by the Gemini integration.

The final script was only done about two days before the presentation. I handwrote it on my tablet, as this method is proven to help with memorization. Later, during the dinner hosted by the organizer for speakers, one of the speakers mentioned how he only managed to complete his slides 10 minutes before his session. If I were in his shoes, I likely would have lost sleep in the days leading up to PyCon.

Autopilot Engaged: The Crisis on Stage

Flustered me on stage (photo courtesy of PyCon MY)

Thinking back, my setup was likely overly complicated and hard to execute. My session was scheduled to be in the bigger auditorium hall. Unfortunately, it clashed with a talk by my project partner from SinarProject. Despite that, many showed up for the talk, and that amplified my uneasiness.

Ideally, I was supposed to have my timer on my phone, so as I went through my script on the tablet I would know how much time was left. Even with that seemingly perfect plan, juggling three devices at a time was almost impossible. I ended up forgetting to start the timer, and forgot about referring to the handwritten script, and just autopiloted through the presentation.

While I didn’t really keep track, I likely missed out a rough guesstimate of 40% of the content. On the other hand, I did put in effort to introduce only one idea in each slide so hopefully the core ideas are presented in a manageable pace.

If you are reading this and you attended the talk, feel free to reach out by commenting here for clarification.

The Return: Why Just Showing Up Was the Answer

If I can, you can too (photo courtesy of PyCon MY)

Attending PyCon as a speaker is indeed an interesting experience. While this is considered a smaller conference (my previous ones were PyCon APAC in KL and EuroPython 2022 in Dublin, Ireland), I was considerably more stressed this time having to deliver a talk.

Siang attending EuroPython 2022 - LaLoka Labs

Until I am done with the talk, that is.

One of the reasons I like attending a conference or meetup is getting to know what people are working on. Apparently, quite a number of speakers are already adopting AsyncIO in their work, which made me wonder if my talk really mattered. This revelation does make me wonder about the people who showed up at my session. Were they just curious about the topic, or wanted to discover new use cases, or simply just there to validate their adoption to the feature?

Perhaps the choice of topic is a major consideration; however, the willingness to put oneself out there matters more. I can now appreciate the support my mentor offered when I was complaining about Rust adoption at work: he pushed me into organizing meetups. Sometimes, the act of standing out and yelling is what it takes to attract more feedback. In this case, speaking in a conference calls back to the problem I had when I started writing about AsyncIO — the constant need for shared knowledge on the topic.

Thanks again for joining me in this wonderful journey. I would like to take the opportunity to express my gratitude towards the organizers for hosting me, and putting in effort to maintain the friendly atmosphere. I shall write again, soon.

This article was developed with the editorial assistance of a large language model, Gemini. The model served as a linguistic and structural advisor, helping the author refine the grammar, correct flow issues, and polish the final prose. The core content, personal voice, and experiences remain the exclusive work of the author, Kitfucoda.