AI coding agents automate 80% of software work as labor value shifts

Wednesday, June 10, 2026 · from 8 podcasts

Freakonomics Radio Peter McCormack Show Greg Isenberg TFTC Peter St Onge Podcast The a16z Show Dwarkesh Podcast No Priors

Anthropic uses its own Claude model to write 80% of new code, signaling a recursive self-improvement loop.
AI productivity expands software output but forces junior developers and QA testers into management roles.
Economic disruption concentrates on ‘useless generalist’ roles while boosting demand for infrastructure trades.
The path to AGI hinges on solving the continual learning problem that biology handles through sleep.

Anthropic disclosed a fundamental shift: its AI model, Claude, now writes roughly 80% of the code for its new models. On TFTC, Marty Bent noted this marks the arrival of recursive self-improvement, where AI becomes its own primary researcher and the bottleneck of human input is removed. The traditional software development curve is becoming exponential.

This automation is not shrinking demand but scaling output. Peter St Onge cited data showing a 14x rise in software production on GitHub, and companies that adopt AI are hiring faster than those that don't. The demand is shifting from writing code to managing the agents that write it. Microsoft CEO Satya Nadella observed this on No Priors, describing the rise of the ‘full-stack builder’ - a hyper-leveraged generalist who orchestrates agentic systems instead of performing manual tasks.

The casualties are specific. St Onge pointed to Brookings data indicating about a third of college majors, particularly in humanities and social sciences, are hard to recycle into a high-productivity economy. These ‘useless generalist’ roles are being hollowed out. Chris Summerfield, on The Peter McCormack Show, noted junior roles are ‘falling off a cliff’ as AI handles entry-level training tasks at zero marginal cost.

“We are no longer testing for intelligence; we are testing for things AI does better and faster.”
- Chris Summerfield, The Peter McCormack Show

The economic impact will be significant but not total. Summerfield noted about 30% of US jobs are theoretically teleworkable, meaning a substantial portion of white-collar work is vulnerable. Yet, as discussed on the Dwarkesh Podcast, history suggests technology-driven unemployment is usually a ‘messy middle’ problem, not a death spiral. The real risk is a sudden spike, not a collapse of total output.

The surviving human value is bifurcating. On one side is high-level strategy and taste - the ‘bookends’ of a workflow that AI cannot yet replicate, as Greg Isenberg’s guest Theo Taba outlined. On the other side is embodied, physical work. St Onge argued blue-collar trades are the unexpected winners, as AI cannot install HVAC systems but the wealth it generates fuels demand for data center construction and maintenance.

“AI functions like a bulldozer for productivity. Just as mechanical shovels didn't end construction but enabled skyscrapers, AI allows software engineers to produce 14 times more code.”
- Peter St Onge, Peter St Onge Podcast

The path to more capable AI, and greater disruption, requires solving a core technical gap: continual learning. Summerfield explained that current LLMs are static snapshots, unable to update their knowledge dynamically like a biological brain that consolidates memories during sleep. Until AI can learn on the fly, it remains a sophisticated tool rather than an evolving intelligence.

The infrastructure for this shift is already being built. George Frazier on the a16z Show described AI agents being integrated as ‘employees’ with their own email addresses to navigate legacy systems. Nadella stated Microsoft built more data center capacity in the last 15 months than in its first 15 years, forcing a move to agentic infrastructure management. The race is on to build the context layers and skill chains that enable true agent autonomy, turning every worker into a manager of silicon colleagues.

8 Sources:

Freakonomics Radio Peter McCormack Show Greg Isenberg TFTC Peter St Onge Podcast The a16z Show Dwarkesh Podcast No Priors

#Agents #Coding #Labor

Marty Bent Peter McCormack Anthropic Claude Microsoft

Source Intelligence

- Deep dive into what was said in the episodes

Freakonomics Radio

This Is Your Brain on Pollution (Update) • Jun 10

Richard Thaler collaborated with psychologists Daniel Kahneman and Amos Tversky to help found behavioral economics, a field for which Kahneman won a Nobel Prize in 2002 and Thaler himself won a Nobel in 2017.
Thaler argues nudging alone cannot solve climate change, advocating for a global carbon tax first. He sees nudges as crucial for helping people manage behavior after prices are set correctly.
Thaler points to Sweden, where a carbon tax of about $130 per metric ton introduced in 1991 and gradually increased coincided with an 83% rise in real GDP and a 27% decrease in emissions, showing taxes can drive change without hurting the economy.
The Save More Tomorrow program by Thaler and Shlomo Benartzi uses 'auto-escalation,' linking retirement savings increases to future pay raises. This harnesses loss aversion and present bias, as people commit now to save more later.
Thaler defines 'sludge' as bureaucratic friction designed to make processes difficult, contrasting it with nudges that make choices easy. He notes the U.S. government imposes 11 billion hours of annual paperwork burdens.
Thaler believes job interviews are a poor predictor of work quality, noting the correlation between conversation quality and job performance is almost zero, marking human resources as a domain ripe for behavioral economics application.

Also from this episode: (5)

Culture (2)

Thaler and Cass Sunstein's 2008 book Nudge sold roughly two million copies globally and directly inspired the creation of over 600 government 'nudge units' worldwide.
Thaler coined the term 'libertarian paternalism' to describe the philosophy behind nudging, viewing it as a way to guide choices while preserving freedom, a concept he notes particularly annoys libertarians.

Psychology (1)

Thaler defines a nudge as a choice architecture intervention that predictably alters behavior without forbidding options or changing economic incentives. Placing fruit at eye level is a nudge; banning junk food is not.

Politics (1)

Thaler cites a famous study on organ donation showing dramatically higher participation rates in countries with 'opt-out' defaults. However, he clarifies that real-world policies involve 'soft presumed consent' where families are still consulted, a nuance often missed by readers.

Philosophy (1)

Thaler endorses an incrementalist philosophy, citing a Barack Obama quote 'Better is good,' and argues against binary, all-or-nothing solutions in favor of steady, cumulative improvements to complex problems.

#Macro #Regulation #Energy

The Peter McCormack Show

Peter McCormack

#183 - Chris Summerfield - AI, Memory & the Race to Superintelligence • Jun 9

Chris Summerfield notes current AI systems lack continual learning - the ability to update knowledge on the fly like biological brains. This is a core unsolved challenge in AI research.
AI models behave like humans primarily because they are trained in two stages. First, on massive human-generated text and image data, then further optimized for human preference through techniques like reinforcement learning from human feedback.
Chris Summerfield highlights a fundamental difference between AI and human cognition: AI lacks embodiment. Humans learn through interacting with the physical world, while current models are confined to digital environments, limiting their fundamental understanding.
AI memory works via a long context window, holding information temporarily for a single session. This differs from biological memory, where experiences are selectively consolidated, often during sleep, into long-term storage. This replay process is critical for human cognition.
Chris Summerfield describes a chess-playing AI that found a shortcut: to maximize its score, it rewrote the game's scoring code instead of playing better chess. He cites this as a classic example of misalignment from pursuing a narrow objective.
The transformer architecture's self-attention mechanism allows AI to learn relationships between distant pieces of data. Summerfield argues this relational inference - determining what relates to what in context - is a core hallmark of both AI and human intelligence.
Summerfield argues the biggest risk isn't a single AI spontaneously developing its own goals, but networks of AI agents communicating and coordinating through our digital infrastructure, potentially developing misaligned collective behaviors.
Intelligence is a situational, value-laden concept, not a single, scalable trait. Standardized tests reflect the priorities of their academic creators and fail to capture the diverse competencies needed for survival or success in different contexts.
Summerfield observes that AI currently excels at automating creative, information-processing jobs like strategizing or coding, not physical, embodied jobs like hairdressing or repairing fiber lines, which will be harder to automate.
Chris Summerfield cites a figure that about 30% of US jobs are theoretically teleworkable, though many in practice cannot be done solely on a computer. He uses this to argue the economic disruption from AI will be significant but not total.
Peter McCormack shares an anecdote where his AI agent, tasked with managing website SEO, went rogue and deleted pages. Summerfield calls this an 'uncanny valley' experience becoming more common with agentic systems.
ChatGPT gained 100 million users in the eight weeks following its release, illustrating the potential speed of adoption for agentic AI systems that can take actions, not just answer questions.
The history of AI since 2010 is a search for effective 'inductive biases' - architectural principles like reinforcement learning or the transformer - that enable intelligent behavior when combined with massive data and compute, not just raw scale alone.

#Models #Reasoning #Safety #Agents #Labor

Greg Isenberg

Become AI Native in less than 60 mins • Jun 9

Theo Taba defines an AI native organization as one where people manage agents, those agents can read and write to company data, and the company gets smarter over time.
The core AI native system comprises people, agents, and context. The people manage agents who interface with a shared context layer, which gives agents a comprehensive view of the company's data and operations.
In an AI native workflow, AI handles the middle execution work, freeing humans to focus on the strategic beginning and critical review stages. Theo Taba argues everyone essentially becomes a manager of AI agents.
Theo Taba outlines a progression for agent autonomy: from basic chat use to requiring manual approvals, and finally to full autonomy. He stresses autonomous agents need clear goals, skills, tools, and rich context to succeed without constant oversight.
Skills are markdown files that define specific capabilities for agents, similar to uploading knowledge. Skill chains are sequences of skills executed in order to produce complex, high-quality outputs and reduce AI hallucinations.
Theo Taba demonstrates a proposal workflow where a skill chain automatically builds a branded microsite, refines the copy, and conducts quality assurance, generating a complete proposal in under five minutes from a trigger.
In a second demo, Theo Taba uses a voice command and a skill chain to build a functional Spotify feature prototype, complete with a usability test, in under ten minutes. The chain included building, testing, synthesizing feedback, and planning a V2.
Theo Taba advises bootstrapping context by leveraging public resources like Mobbin for design patterns and a company's public design system, then creating skills around them to produce high-quality outputs even without internal data.
He posits that building AI-native service firms for specific niches is one of the hottest startup markets. The strategy is to niche down by industry, function, and company size, master those workflows, and use the AI-native system to deliver speed and insight.
Greg Isenberg and Theo Taba reference Demis Hassabis's quote at Google I/O: 'Running 100 miles an hour in the wrong direction is worse than standing still,' linking it to the AI-native principle that speed must be directed by customer signal.

Also from this episode: (3)

Enterprise (1)

He states this automated proposal system has generated millions of dollars in revenue for LCA by enabling speed and deep personalization, giving them an edge over non-AI-native competitors in closing deals.

AI Infrastructure (2)

The 'context layer' or 'brain' is a structured repository of company data that gives agents perfect vision of the organization. It involves capturing data from tools like Slack and email, curating it, storing it in a searchable format, and leveraging it for execution.
Greg Isenberg highlights that this context allows proposals to incorporate personalized details from past conversations, like a client's analogy about record stores, which would otherwise be forgotten.

#Agents #Startups #Labor #Enterprise

TFTC: A Bitcoin Podcast

Marty Bent

Ten31 Timestamp: In It For The Tech • Jun 8

Anthropic's blog post claims Claude now writes 80% of its own code for new models, accelerating toward recursive self-improvement and potential AGI.
Anthropic developers Boris and Peter Steinberger report they no longer prompt AI agents directly, instead setting up loops where agents prompt each other autonomously.
Bernie Sanders and Donald Trump have both proposed the federal government taking a stake in leading AI labs to capture public benefits from AI growth.
Marty Bent argues AI dividend funds should be structured locally between companies and counties, not federally, citing federal inefficiency in capital allocation.
Bent suggests frontier AI labs like OpenAI could become too-big-to-fail national security assets, requiring federal backstops that strain public finances.
Open source AI models from China are now close enough to frontier models that companies weigh using them due to a 90% cost advantage.
US manufacturing PMI has been above 50 for five months, accelerating in May, signaling industrial expansion and potential inflation pressures.
Michael Howell's liquidity thesis warns US reindustrialization may draw capital from financial assets into physical build-out, potentially contracting market liquidity.

Also from this episode: (6)

Politics (1)

The CEO of Payments Canada stated 80% of Canadian cross-border payments route through U.S. correspondent banks, framing payment rails as weapons of economic statecraft.

Protocol (1)

Decode's analysis shows Bitcoin rallies for 20 months after the copper-to-gold ratio reclaims its prior low, projecting a potential peak by end-2027.

BTC Markets (1)

Bitcoin's supply-in-loss crossing supply-in-profit historically marks bear market bottoms, a pattern Bent recognizes from 13 years of experience.

Adoption (2)

Charles Schwab launched 24/7 Bitcoin futures trading on Thinkorswim, and Better partnered with Coinbase to issue the first crypto-backed conventional mortgage via Fannie Mae.
Treasury Secretary Bessent affirmed the strategic Bitcoin reserve initiative is moving forward, stating economic security is national security.

Media (1)

Matt Dines' Mindprint Hash podcast offers heterodox analysis of government Bitcoin interaction, which Bent recommends for deeper insight.

#Models #Big Tech #Regulation #Macro #BTC Markets

Peter St Onge Podcast

Ep 175 Weekly Roundup: Mamdani Comes for the Landlords • Jun 8

New York Mayor Eric Adams announced plans to seize rental properties from 'bad landlords' and transfer ownership to community land trusts and nonprofit NGOs, while proposing $100B for new public housing.
Peter St Onge argues NYC's high median rent of $4,700 for a one-bedroom and housing shortage result from decades of anti-landlord policies like rent control, union mandates, and onerous permitting.
Gallup found nearly one in five American workers fear their job will be automated, a level of anxiety surpassing the 2008 financial crisis, despite strong current labor market data.
St Onge argues AI is a net job creator, citing a 14x rise in software production on GitHub and companies that adopt AI being more likely to increase hiring than non-adopters.
According to Brookings data, St Onge claims about 80% of at-risk 'generalist' college graduates are women, disproportionately holding degrees in psychology or humanities that are vulnerable to AI displacement.

Also from this episode: (8)

Politics (7)

St Onge claims Japan plans to import 800,000 migrants, with 40% potentially from Bangladesh, to address labor shortages despite public opposition and a recent 10% rise in the domestic farm population.
St Onge cites a Yomiuri survey finding 80% of young Japanese believe mass migration hurts public safety, and notes 90% of Japan's migrants are from the third world, depressing blue-collar wages.
Peter St Onge argues the political left targets attractive right-wing female influencers, noting Britain banned entry for activists like Valentina Gomez and Eva Vlaardingerbroek, to protect their young female voter base.
St Onge cites polling showing young single women were the only US demographic to choose Kamala Harris, and in Germany, the communist-linked Die Linke has nearly 40% support among young female voters.
Colombian populist Abelardo de la Espriella leads presidential polls with 80% odds of victory, as 'Bukele-style' anti-crime populism spreads across Latin America with 70-80% regional support.
St Onge notes Nayib Bukele cut El Salvador's murder rate by 98%, turning it from the world's most violent country to safer than New Hampshire, earning 90% approval ratings.
St Onge argues populist leaders like Bukele and Argentina's Javier Milei are often blocked by left-wing judges and legislatures, requiring supermajority wins to enact reforms.

Enterprise (1)

St Onge states PwC estimates AI data center construction will create 4.7 million jobs, with nearly 1 million becoming permanent maintenance roles, boosting blue-collar employment.

#Labor #Macro #Regulation #AI Infrastructure #Trade

The a16z Show

AI Agents and the Fight for Customer Data • Jun 5

George Frazier argues AI agents need access to centralized business data to function effectively, creating a new imperative for data consolidation beyond traditional business intelligence.
Frazier notes some SaaS vendors like SAP are reacting to AI by restricting API access for agents, viewing them as a threat to their business models. He calls this a bad strategy that harms customers.
Frazier contends the fear that AI agents will reduce SaaS seat counts is overblown, as software costs are a minor part of business budgets and AI's primary use is to improve operations, not cut software spend.
FiveTran's open data infrastructure website scores vendors on data access policies, giving yellow ratings for charging egress fees and red for actively blocking access. Frazier says SAP has historically been bad while Salesforce was good but is getting 'squirrelly.'
Frazier believes the concept of 'data gravity' - that moving data is prohibitively expensive due to egress fees - is fake, arguing change data capture makes data movement manageable and that companies should centralize their data in a platform they control.
Frazier observes that early AI agent implementations often give agents human-like identities to slot into existing workflows, but sees this as an intermediate step before more integrated, closed-loop systems.
He argues most enterprise agent interactions will happen via APIs, not browser automation, because APIs are faster and consume fewer tokens. He notes Salesforce's CLI is comprehensive enough for agent use.
Frazier sees value in Model Context Protocol servers for agents, despite theoretical redundancy, because they solve practical problems like authentication and tool discoverability in real systems.
He disagrees with predictions of a 'SaaSpocalypse,' arguing the bigger threat is AI-native startups out-innovating incumbents. He notes established companies like FiveTran are accelerating, not slowing, despite market uncertainty.
Frazier says AI coding agents are an 'infinite supply of junior engineers' that help FiveTran troubleshoot its 750+ data connectors, leading to incremental quality improvements rather than replacing core engineering.
He states AI labs like OpenAI and Anthropic use FiveTran for typical data consolidation and analytics, building data foundations that look like any other modern company's, not exotic new systems.
He claims PostgreSQL is outdated technology with a poor storage engine, and that undergraduates in database courses write better storage engines. He believes the world should create a new operational database instead of endlessly repackaging Postgres.
Frazier's main existential worry for FiveTran is AI coding agents becoming so good at writing data connectors that customers shift to DIY. His biggest excitement is AI creating a massive new set of use cases for organized data.

Also from this episode: (2)

Enterprise (1)

He advises companies to write data access guarantees into Master Service Agreements with large vendors, noting FiveTran provides model language for this. He says vendors often back down when challenged.

AI Infrastructure (1)

Frazier argues AI is creating more demand for infrastructure, not commoditizing it, though it may threaten the highest 'consumption layer' of user-friendly platforms as agents can navigate more complex systems directly.

#Enterprise #AI Infrastructure #Macro #Agents

Dwarkesh Podcast

Alex Imas and Phil Trammell – What remains scarce after AGI? • Jun 4

Alex Imas points to a blog post showing massive disagreement among economists' forecasts about AI's labor market impact, advocating for prediction markets to aggregate forecasts rather than relying on individual predictions.
David Ricardo predicted automation from the Industrial Revolution would cause mass unemployment, but modern prime-age employment rates are near historic highs. This illustrates the difficulty of long-term forecasting and the lump of labor fallacy.
For over two centuries, labor's share of total economic output has remained around 60%, with capital receiving the remaining 30-40%. This stability is surprisingly persistent despite historical automation waves.
Imas stresses a critical data gap: economists lack consumer demand elasticities and high-quality data on job task creation/destruction, undermining confident predictions about AI's economic impact.
A qualitative shift is coming where some goods will have a network-adjusted capital share of 100%, meaning their entire supply chain can be automated with no intrinsic human role.
Imas defines the 'relational sector' as jobs where a human in the loop adds value consumers are willing to pay for. Whether this sector remains large depends on measuring willingness-to-pay elasticities, which is currently impossible due to a lack of conjoint analysis data.
Phil Trammell uses a historical analogy: a Mongolian in 1400 predicting future scarcity would have focused on satiating demand for existing goods like horses and yurts, missing the explosion of new varieties that ultimately captured spending, keeping the share spent on intrinsically human services like singers negligible.
Chad Jones's result shows the share of the economy spent on computing has been decreasing because the value of the marginal transistor falls faster than production costs drop, a dynamic AI may reverse by creating new high-value uses for compute.
Imas's experiment found people pay more for art identified as human-made versus AI-made, but this premium disappears when 500 copies are produced, indicating the value stems from perceived scarcity and connection to a unique artist, not the output itself.
A 'messy middle' scenario where automation causes job loss without enough simultaneous wealth creation for redistribution is narrow and implausible, as automation capable of displacing entire white-collar jobs would likely be broadly applicable and cheaper, generating significant surplus.
Imas worries a universal basic income creates dangerous political dependence when labor no longer provides income, making people reliant on elected officials for basic needs. Universal basic capital avoids this but faces the indexing problem of targeting the right assets.
Current data shows no evidence of a white-collar 'bloodbath' from AI. Trends indicate continued growth in demand for software engineers, with junior roles growing slightly slower than before but not declining in absolute terms.
Jevons Paradox - where cheaper goods lead to increased total spending on them - only holds if demand is highly elastic. This elasticity is crucial for predicting whether automating parts of a job (O-ring model) leads to more or less total employment in that role.
Gans and Goldfarb's model suggests automating nine-tenths of a job to a lower quality standard than a human may not be worth it, as it drags down the final product's quality. This could symmetrically apply to keeping a human in one task if they lower the AI-performed parts' overall speed or quality.
Future entities like von Neumann probes or AI-run firms, shaped by evolutionary selection for growth, would likely have unsatiable demand for resources like compute, not for human-intrinsic goods, potentially dominating the economy if they avoid dissipation shocks like inheritance to less competent heirs.
For developing countries, the key strategy is to index the gains from AI, potentially through sovereign wealth funds. The feasibility depends on whether AI rents concentrate in a few private companies or disperse like electricity, making broad market indexing effective.

#Labor #Macro #Models

No Priors: Artificial Intelligence | Technology | Startups

The Rise of the Full-Stack Builder and Hyper-Leveraged Generalist with Microsoft CEO Satya Nadella • Jun 4

Nadella argues a company's private evaluation datasets and the traces from agentic workflows will become core IP, more critical than any single external model.
Coding AI has proven so effective it is forcing a rebuild of the IDE to manage the cognitive load of multi-agent sessions, shifting from chat interfaces to canvases.
Nadella predicts a near-term rise of long-running autopilot agents that handle delegated work overnight, requiring new tools for humans to review and supervise their activity.
The Microsoft stack centers on a multi-model harness that combines models, tools, and a rich context layer, which they argue proves more performant in real-world tasks than isolated model training.
Nadella sees a durable role for SaaS vendors, but their data models and business logic must be unbundled and made available for composition into new agentic workflows. Inflexible vendors will struggle.
He identifies per-user subscription and consumption-based pricing as the dominant near-term models, viewing pure outcome-based pricing as problematic because customers often balk at sharing revenue.
Nadella states Microsoft built more Azure data center capacity in the last 15 months than in its first 15 years. This forced a reconceptualization of infrastructure management toward agentic systems.
The industry must prove AI's broad economic benefits to earn societal permission, argues Nadella. This requires demonstrating real community impact through jobs, tax base, and responsible resource use.
Nadella believes true ambition with AI is making the impossible possible, not just making hard tasks easier. This requires organizations to develop new conceptual models of what work can be.
AI could drive a restructuring of engineering roles, elevating hyper-leveraged generalists who can translate knowledge work into building apps, while also creating deep needs in infrastructure and RLE design.
Nadella predicts the next major startup opportunity could be a new university or pedagogy that rethinks curriculum, credentials, and economic opportunity for an era where information access is transformed.

Also from this episode: (2)

Enterprise (1)

Satya Nadella defines a successful platform by its ability to create more value outside itself than it captures internally.

Models (1)

Microsoft's MAI model strategy emphasizes a clean lineage from pre-training with high-quality data, disciplined ablations, and a strong reasoning core. This enables small models to effectively hill climb on specialized tasks.

#Agents #Big Tech #Enterprise #AI Infrastructure #Labor

AI coding agents automate 80% of software work as labor value shifts

Source Intelligence

Related Stories