Narrative · ↗ rising Software factory: agents now write code All narratives · 19 supporting stories · 1 from Hungary
„At leading companies, AI already writes 80–90% of code, and software factories built from agent swarms are replacing traditional development teams.”
Counter-narrative According to Jellyfish data covering 1,300 companies, agents affect 27.7% of pull requests at the median firm. However, in the software factory model where agents open PRs, this figure is 0.0% for 90% of companies. A Hungarian report states Meta's experiment to replace programmers with AI failed. This view is primarily supported by independent measurements.
Carriers Statements by Anthropic, OpenAI and GitHub/Microsoft themselves, Factory, Cognition (Devin), Cognizant, Gergely Orosz (The Pragmatic Engineer), Venture capital (Factory investors)
Analysis According to the narrative, development has shifted to industrial scale. This is indicated by agent fleets, OpenAI's new Git backend (Pando, after a 2442% increase in commits), Anthropic's rebuilding coding infrastructure, and Czech companies banning manual coding. The vast majority of these figures are provided by those selling agents or agent platforms. Factory's $5 billion valuation is built specifically on the software factory concept. The AI lab as its own first customer is not a representative sample due to its unique monorepo, culture, and budget. Independent Jellyfish data (0.0% in nine-tenths of companies) and Meta's failure argue against generalizing the narrative. This is one of the most claim-driven stories of the period.
Evidence: weak — The narrative rests almost entirely on vendors' and tool makers' own statements: 90%, 80%, 800 thousand lines, 1211 PRs in a single day, 37% savings. Independent Jellyfish measurement shows broad but partial agent use, and finds the software-factory model rare. The results of METR's independent evaluation cannot be determined from the source.
What's at stake: Whether evaluations based on manufacturer self-declarations and work organization decisions can withstand independent measurement.
Narrative sub-threads The narrative splits into 5 branches; a story may belong to several branches, and some may not appear in the main narrative's list.
▲ at peak · evidence: weak Self-reported rates of AI-written code According to their own statements, frontier companies already have 75–90% of their code written by AI. Anthropic states that 90% of the Claude Code code is AI-written, while a Cybernews report says 80% of the company's code is written by Claude. Reports indicate that nearly 75% of Google's (new) code is AI-generated.
This branch relies on large tech companies' own ratios, not on tools or organizational models. Reports accompanying Google data indicate that despite high ratios, meaningful acceleration is lacking and developer distrust is rising. METR published an independent assessment of Anthropic's R&D acceleration, but its results are not clear from the source.
4 stories · 0 of them Hungarian → lasting · evidence: moderate Large-scale rewrites performed by agents Entire systems are being rewritten or rebuilt with agents. Reports state that 800,000 lines of Rust code for the Copilot runtime environment were written by AI for 120,000 dollars. A new kernel's approximately 139,000 lines were written by a Codex agent, and the kernel was completed in 11 days. Asana replaced an outdated test system with Codex in two weeks for roughly 12,000 dollars.
Here the evidence is not a ratio but a single migration and rewrite project, quantified in lines of code, cost and time. The performance claims (16-fold and 93-fold speedups) are the creators' own statements.
6 stories · 1 of them Hungarian ↗ rising · evidence: moderate Infrastructure for the agent code flood Due to the volume of commits and pull requests generated by agents, new version control backends and capacity are being built. Commit counts in OpenAI's monorepo rose by 2442 percent, prompting the company to build the Pando Git backend. A summary states GitHub is adding 4 million CPU cores and 160 PB of storage. In August, agent pull requests on GitHub exceeded human ones.
This branch builds on the side effect of volume: it shows the shift through the load on Git hosting and repositories, not through code authorship.
6 stories · 0 of them Hungarian ↗ rising · evidence: moderate The software factory model and its market Platforms organizing agent swarms as teams, and companies built on the 'software factory' concept, are growing: Factory's valuation tripled to $5 billion. However, according to Jellyfish's data from 1,300 companies, the factory model, in which agents open PRs, stands at 0.0% at nine in ten companies.
This segment covers the organizational model and its productization (Kiro Crew, Warp Factories, Factory evaluation, AMD swarms, the Codex at OpenAI). It also presents measurements contradicting widespread adoption.
8 stories · 0 of them Hungarian → lasting · evidence: weak Developers replaced by agents Some companies would replace programmers or hiring with AI. In Beijing, 160 programmers were laid off two weeks after the boss asked about AI replacement; the source does not confirm causality. Czech IT firms banned manual coding, and there were layoffs. According to a Cognizant–Cognition statement, 37% net savings are achievable with Devin. At Anthropic, the share of junior technical hires fell to 9%.
This strand rests on workforce-side consequences: cases of dismissal and replacement, plus one failure (Meta), not technical or volume metrics.
5 stories · 1 of them Hungarian Sub-thread stories Self-reported rates of AI-written code Carriers: Anthropic, Google, METR
Oct 16, 2026 Nearly 75% of Google's code is now written by AI, but speed-up lags CTO The share of AI-written code is not the same as delivery speed; measure lead time.tech lead Increase review and testing capacity as well if more code is being produced.Oct 19, 2026 Anthropic: AI wrote 90% of Claude Code's code, with 30,000 changes a day tech lead With a high share of AI-written code, scale review and infrastructure to the volume of changes.Nov 2, 2026 AI writes 75% of Google's new code as developer distrust rises tech lead As the share of AI-generated code grows, code review culture and test coverage need to be strengthened in advance.senior developer Review has become your main job, so it is worth developing methods for checking AI-generated code.Nov 5, 2026 Claude writes 80% of Anthropic's code; METR publishes independent assessment of its AI R&D acceleration tech lead If your own team also relies on AI-written code, set up independent review and control processes in advance.Large-scale rewrites performed by agents Carriers: GitHub/Microsoft, OpenAI Codex, founder of Rocky Linux, Asana, Mark Kern
Aug 18, 2026 Asana replaces a legacy test system in two weeks with Codex, at about $12,000 QA tester Test-system migration can be accelerated with agents, but coverage and correctness must be confirmed by human review.tech lead Long, mechanical migrations are good candidates for an agent pilot, with cost and deadline estimates.Sep 12, 2026 Former Blizzard developer replaces Unreal Engine with AI-built custom engine, claiming 93x speedup senior developer The claim is self-reported and has no independent measurement, so treat it with caution, but it shows the direction of large AI-assisted rewrites.Sep 16, 2026 GitHub moves Copilot runtime to Rust, with AI agents writing 800,000 lines of code senior developer The case is a practical example of a large-scale language migration carried out with agents.tech lead Migration plans should weigh agent-based code rewriting and the test coverage it requires.CTO A Rust migration with AI is a possible way to reduce technical debt and is worth trying as a pilot.Sep 21, 2026 Microsoft used AI to rewrite the Copilot runtime in Rust for $120,000 CTO The cost and performance figures can support business cases for AI-assisted migration, but the source should be verified.tech lead Consider which legacy components could be earmarked for an AI-driven rewrite and how the result can be measured.senior developer Expect review and testing to be the main work in a large-scale rewrite.Oct 9, 2026 GitHub migrates Copilot runtime to Rust with AI-assisted rewrite senior developer In a large rewrite project, AI assistance provides a real, vendor-level example.tech lead In migration plans, it is worth involving AI help, but test coverage should be secured first.Oct 26, 2026 Rocky Linux creator builds new non-Linux kernel in 11 days using Codex senior developer The example shows that system-level code can also be produced with an agent, but verification remains the human's job.Infrastructure for the agent code flood Carriers: OpenAI, Anthropic, GitHub, Cursor, Harness
Aug 11, 2026 Cursor releases Origin, a Git host for fleets of AI agents tech lead Assess whether agent fleets need a separate Git host or whether your current platform is enough.Sep 7, 2026 Harness launches AI code review and rebuilds its Git repository for agent traffic tech lead AI code review can help with the review bottleneck if false alarms are measured.Oct 5, 2026 Agent pull requests overtook human ones on GitHub in August as Copilot moves to usage-based pricing tech lead Given the share of agent PRs, review capacity and PR rules need to be redesigned.CTO With usage-based pricing, Copilot costs depend on load, so a spending cap should be set.senior developer More agent PRs need to be reviewed, so review priorities should be set down in writing.Oct 12, 2026 Anthropic built its own source control for AI-scale code volume, but it is not available CTO AI-generated code volume carries an infrastructure cost; plan for CI and storage capacity.tech lead Set limits on the number of branches and PRs that agents open.Oct 28, 2026 GitHub adds 4 million CPU cores and 160 PB of storage as AI code surge strains capacity CTO Agent-generated traffic also strains platform capacity, so it is worth planning a fallback.Nov 4, 2026 OpenAI builds new Git backend Pando after Codex agents drive commit surge CTO Adopting agents can mean investment in storage and CI, not just tool licences.The software factory model and its market Carriers: AWS, Factory, Warp, AMD
Aug 4, 2026 AWS launches Kiro Crew, an autonomous orchestrator for 24/7 coding agents tech lead Decide which tasks can be given to unsupervised agents and where human review remains mandatory.CTO The cost and accountability model for continuously running agents should be settled before deployment.senior developer The centre of gravity of the work shifts toward defining agents' tasks and reviewing their results.Aug 17, 2026 AMD's agent swarms are rewriting its software development workflow, IEEE Spectrum reports senior developer The centre of gravity of the job is shifting toward writing specifications and validating results, so these are the skills to build.tech lead Tasks need to be broken down so that agents can work in parallel, with human review acting as a gate.Aug 18, 2026 Warp launches Warp Factories, bringing the AI-factory model to software development tech lead For workflows handed to agents, define gates and metrics before the team works in a factory model.Aug 31, 2026 AWS open-sources Kiro Crew agent system tech lead Worth trying as a framework for running durable agents across multiple tools.Sep 15, 2026 AI coding agent startup Factory triples valuation to $5 billion with $200 million round CTO The agent-based "software factory" model is a candidate for enterprise pilots.startup founder The funding figure signals the strength of the market for agent-based development tools, but it comes with strong competition.Sep 15, 2026 Gergely Orosz: Codex has "taken over" OpenAI as an agentic software factory takes shape tech lead OpenAI's internal example can be used to design an agent-centred development process.CTO The report is a practical reference for organisational transformation.Oct 12, 2026 Factory AI's valuation triples in five months to $5 billion startup founder In the agent-based developer-tools market, abundant capital means strong competitors; focus on a narrow, defensible niche.CTO The tools market is consolidating quickly, so avoid long, exclusive contracts.Oct 26, 2026 Jellyfish: agents touch 27.7% of pull requests, while "software factories" remain rare tech lead Agents currently assist human work more than they open PRs on their own, so tailor the review process accordingly.CTO Full automation is still rare, so the PR share is a realistic target to measure.Developers replaced by agents Carriers: Cognition/Cognizant, Czech IT companies, Anthropic, Meta
Aug 24, 2026 Beijing developer and 160 colleagues laid off two weeks after boss asked about AI replacing coders junior developer Show measurable value, including with AI tools, and keep up skills beyond your salaried role.senior developer Document your own contribution, which an agent does not replace.Aug 27, 2026 Meta's attempt to replace programmers with AI reported as a failure CTO Run a pilot before any replacement plans, and measure quality and rework needs.tech lead Use it as an argument that replacing headcount with AI carries risk.Sep 23, 2026 Cognizant and Cognition put autonomous AI engineering into production at Odyssey Logistics, citing 37% net savings CTO The vendor figure can be treated as a business case, but only alongside independent validation.tech lead Check which multi-step development tasks can be handed to an agent.Sep 30, 2026 Czech IT firms: programmers banned from writing code, some laid off senior developer Prepare for roles to change: review and specification may become the main task instead of writing code.CTO Companies in the region already have examples of coding bans, so it is worth planning your own policy and retraining.Oct 19, 2026 Anthropic's junior technical hiring fell from 27% to 9% as AI agents take on work junior developer Fewer entry-level routine positions will be available, so checking and testing agents' work may be the distinguishing skill.tech lead Plan deliberately how juniors will learn when agents take over routine tasks.CTO Hiring plans and training budgets should be reconsidered, because a drained junior pipeline is a long-term succession risk.Google search interest Google Trends relative search index over the period studied, weekly average. A narrative and its sub-threads appear in a single comparison: 100 is the group's most-searched day, and the figures can only be compared within the group. The search term was chosen by machine analysis; it measures how much the term is searched, not how widespread the narrative is.
Who is affected and how CTO The 90% figures should not be considered benchmarks. First, measure on your own PR data how many percent are actually affected by an agent (the Jellyfish median is 27.7%), and what the error rate is there. tech lead Large-scale agent execution has significant costs: the Copilot runtime Rust rewrite reportedly cost 120,000 dollars. Therefore, the factory model should only be tested with a budget and a small pilot. senior developer The role shifts toward specification and validation (AMD case). It is worthwhile to develop skills in task description, testable acceptance criteria, and agent output verification. startup founder Investors reward the software factory story (Factory, $5 billion). However, savings figures such as 37% come from vendor statements. It is advisable to collect your own auditable ROI data. Hungary Domestic traces during the period under review are scarce. The only [HU]-candidate story is regional: according to Lupa.cz's report on September 30, several Czech IT companies banned programmers from manual coding due to AI, and some people were dismissed (2026-09-28-20). No domestic measurement exists.
Weekly trend Supporting stories per week
3 Aug 3–9: 0 Aug 10–16: 0 Aug 17–23: 1 Aug 24–30: 2 aug. 31 – Sep 6: 0 Sep 7–13: 2 Sep 14–20: 3 Sep 21–27: 2 szept. 28 – Oct 4: 1 Oct 5–11: 0 Oct 12–18: 1 Oct 19–25: 2 okt. 26 – Nov 1: 2 Nov 2–8: 3 Aug 3–9 Nov 2–8
Turning points Sep 15, 2026 Gergely Orosz: Codex has "taken over" OpenAI as an agentic software factory takes shape Orosz's September 15 analysis showed how Codex became OpenAI's everyday development tool and how an agent-based software factory is being built there. Oct 19, 2026 Anthropic: AI wrote 90% of Claude Code's code, with 30,000 changes a day Anthropic stated that 90% of the Claude Code codebase was written by AI. The number of daily changes exceeded 30,000, prompting the company to rebuild its coding infrastructure. This has become the narrative's key reference figure. Oct 26, 2026 Jellyfish: agents touch 27.7% of pull requests, while "software factories" remain rare According to Jellyfish data, agents affect 27.7% of PRs at the median company, while the software factory model shows 0.0% for nine-tenths of companies. This makes manufacturer figures appear as an exception for a narrow elite. Related narratives in the News magazine Supporting stories Aug 17, 2026 AMD's agent swarms are rewriting its software development workflow, IEEE Spectrum reports senior developer The centre of gravity of the job is shifting toward writing specifications and validating results, so these are the skills to build.tech lead Tasks need to be broken down so that agents can work in parallel, with human review acting as a gate.Aug 24, 2026 Beijing developer and 160 colleagues laid off two weeks after boss asked about AI replacing coders junior developer Show measurable value, including with AI tools, and keep up skills beyond your salaried role.senior developer Document your own contribution, which an agent does not replace.Aug 27, 2026 Meta's attempt to replace programmers with AI reported as a failure CTO Run a pilot before any replacement plans, and measure quality and rework needs.tech lead Use it as an argument that replacing headcount with AI carries risk.Sep 7, 2026 OpenAI says its researchers increasingly rely on coding agents for more of their working hours CTO Agent working time and cost should be planned as a separate cost line.tech lead Rolling out an agent-based workflow requires usage and cost tracking.Sep 7, 2026 Former Lovable founders launch Blaise AI to replace IT consultancies startup founder AI-based providers are targeting the IT consulting market, which means new competitors and suppliers.Sep 15, 2026 AI coding agent startup Factory triples valuation to $5 billion with $200 million round CTO The agent-based "software factory" model is a candidate for enterprise pilots.startup founder The funding figure signals the strength of the market for agent-based development tools, but it comes with strong competition.Sep 15, 2026 Gergely Orosz: Codex has "taken over" OpenAI as an agentic software factory takes shape tech lead OpenAI's internal example can be used to design an agent-centred development process.CTO The report is a practical reference for organisational transformation.Sep 16, 2026 GitHub moves Copilot runtime to Rust, with AI agents writing 800,000 lines of code senior developer The case is a practical example of a large-scale language migration carried out with agents.tech lead Migration plans should weigh agent-based code rewriting and the test coverage it requires.CTO A Rust migration with AI is a possible way to reduce technical debt and is worth trying as a pilot.Sep 21, 2026 Microsoft used AI to rewrite the Copilot runtime in Rust for $120,000 CTO The cost and performance figures can support business cases for AI-assisted migration, but the source should be verified.tech lead Consider which legacy components could be earmarked for an AI-driven rewrite and how the result can be measured.senior developer Expect review and testing to be the main work in a large-scale rewrite.Sep 23, 2026 Cognizant and Cognition put autonomous AI engineering into production at Odyssey Logistics, citing 37% net savings CTO The vendor figure can be treated as a business case, but only alongside independent validation.tech lead Check which multi-step development tasks can be handed to an agent.Sep 30, 2026 Czech IT firms: programmers banned from writing code, some laid off senior developer Prepare for roles to change: review and specification may become the main task instead of writing code.CTO Companies in the region already have examples of coding bans, so it is worth planning your own policy and retraining.Oct 12, 2026 Factory AI's valuation triples in five months to $5 billion startup founder In the agent-based developer-tools market, abundant capital means strong competitors; focus on a narrow, defensible niche.CTO The tools market is consolidating quickly, so avoid long, exclusive contracts.Oct 19, 2026 Anthropic: AI wrote 90% of Claude Code's code, with 30,000 changes a day tech lead With a high share of AI-written code, scale review and infrastructure to the volume of changes.Oct 21, 2026 Recursive launches Flint, a free coding model that runs on laptops senior developer Try it in a local environment to see how usable a laptop-run model is on your own codebase.CTO Locally run models can reduce API costs and the risk of data leakage, so they are worth evaluating.startup founder A free, local model can lower the entry cost of prototype development.Oct 26, 2026 Jellyfish: agents touch 27.7% of pull requests, while "software factories" remain rare tech lead Agents currently assist human work more than they open PRs on their own, so tailor the review process accordingly.CTO Full automation is still rare, so the PR share is a realistic target to measure.Oct 26, 2026 Rocky Linux creator builds new non-Linux kernel in 11 days using Codex senior developer The example shows that system-level code can also be produced with an agent, but verification remains the human's job.Nov 2, 2026 Copilot Cowork generated 1,211 pull requests in a single day tech lead The flood of agent-generated PRs strains the team's review capacity, so plan filters and limits.senior developer Expect to review more machine-made PRs; without review prioritisation this quickly becomes a bottleneck.Nov 4, 2026 OpenAI builds new Git backend Pando after Codex agents drive commit surge CTO Adopting agents can mean investment in storage and CI, not just tool licences.Nov 5, 2026 Claude writes 80% of Anthropic's code; METR publishes independent assessment of its AI R&D acceleration tech lead If your own team also relies on AI-written code, set up independent review and control processes in advance.Machine analysis based on story titles and summaries, with a verification pass. The source of the facts is the article linked at each story.