Skip to sources
Time Machine

Software Developers

Six people did this job in 1945. Today: 1.69 million. every figure cited

Drag the slider to travel 91 years of this work.

2026drag to travel through time
195019752000now
Country
2026
Known today as Software Developer
US Employment
1.69M
BLS OEWS May 2025 (published). Replaced an earlier placeholder ESTIMATE of 1,650,000 once the official release became available, per the curation note that invited the refresh. OEWS is a point-in-time survey snapshot, not a continuous time series; BLS advises against using it for year-over-year trend comparison.
Median Annual Wage
$135,980
≈ $132,494 in 2024 dollars
Each dot is a cited figure over time; the dotted line only links them (values between aren't measured). Hollow dots are estimates.
Beat · 2025

Employment of US computer programmers (the sibling SOC 15-1251, kept separate from software developers) falls to its lowest level since 1980 — about half the 2000 peak — with a 27.5% drop in the 12-month average since early 2023, correlating with ChatGPT's release.

Tools of the era

The tools that defined the work

Select an era to see how it reshaped the work.

  • ENIAC and plug-board programming

    No programming languages existed. Programs were specified by physically rewiring patch panels and setting thousands of switches by hand. Programming was treated as clerical work and largely assigned to women; the engineers who designed the machine considered hardware (not software) to be the intellectual achievement.

    Effect on the work

    Cohort: 6 ENIAC programmers in 1945. The broader US Army "computers" pool — ~200 women doing manual calculation — was the recruiting feeder.

    Work toolChanging equipment
  • FORTRAN, LISP, COBOL, ALGOL

    First-generation high-level languages decoupled programming from machine wiring. FORTRAN (IBM, 1957) made scientific computation a writeable activity; LISP (McCarthy, 1958) seeded the AI tradition; COBOL (1960) put programming inside enterprise back-offices. Headcount grew from ~15K (1957) toward >100K within a decade.

    Mainframe processingComputerized records
  • Structured programming + UNIX + C

    Dijkstra's 1968 "GOTO considered harmful" letter triggered a decade of methodological re-tooling. UNIX (Thompson + Ritchie, Bell Labs, 1969) and C (Ritchie, 1972) made portable systems software possible and seeded the open-source tradition. The job title "software engineer" emerged — formalized at the 1968 NATO conference.

    Work toolChanging equipment
  • IBM PC + Borland Turbo Pascal

    The IBM PC (1981) and the $49.95 Turbo Pascal compiler (Borland, 1983) democratized programming — a single developer could buy a development environment for the price of a textbook. Programming moved out of computer centers and onto individual desks. By 1990 BLS counted ~565,000 computer programmers in the US, ~3.5× the 1980 number.

    Work toolChanging equipment
  • Visual Basic + RAD + Windows business apps

    Microsoft Visual Basic (1991) created an entire occupational sub-category — "VB programmer" — focused on internal business applications. Rapid Application Development tools (VB, Delphi, PowerBuilder) let business analysts build their own software. This is also when "the programmer who isn't a computer scientist" became the modal hire.

    Work toolChanging equipment
  • Open source — Linux, Apache, Perl, MySQL

    The LAMP stack (Linux 1991, Apache 1995, MySQL 1995, Perl 1987→1994) made it possible to build commercial software entirely on free tools. Open source flipped the economics of software development; by the early 2000s no startup paid for a database license, a web server, or an OS.

    Work toolChanging equipment
  • Web 1.0 — Netscape, Java, JavaScript

    The Netscape Navigator browser (1994) and the simultaneous arrival of Java + JavaScript (1995) created an entirely new programming target: the browser. Demand for web developers exploded; "webmaster" and "web programmer" were briefly distinct census-tracked occupations. The dot-com peak in 2000 saw US programmer headcount cross 700,000.

    Work toolChanging equipment
  • Agile + Scrum + XP

    The Agile Manifesto (signed February 2001 by seventeen developers at Snowbird, Utah) replaced the waterfall planning culture that had dominated since the 1968 NATO conference. Scrum (1995) and Extreme Programming (1996) provided the practice scaffolding. Developer time shifted from up-front specification toward iterative iteration in two-week cycles.

    Work toolChanging equipment
  • AWS S3 + EC2 — public cloud

    Amazon S3 (March 14, 2006) and EC2 (August 2006) made it possible to ship a production service without buying a single server. By 2010 every Amazon.com retail site had migrated to AWS; by 2012 Netflix had publicly committed 100% of infrastructure to AWS. The "ops" half of the developer job mostly disappeared into AWS bills.

    Work toolChanging equipment
  • GitHub + distributed version control

    GitHub launched April 2008. Linus Torvalds's git (2005) had already replaced CVS/SVN in serious open-source projects; GitHub built a social layer on top and within a decade had >100M users. Open-source contribution became a public résumé. Hiring funnels reorganized around public commit history.

    Work toolChanging equipment
  • DevOps + Docker + Kubernetes

    The first DevOpsDays conference (Ghent, October 2009, founded by Patrick Debois) named a practice that had been emerging at Flickr, Etsy, and Amazon. Docker (2013) and Kubernetes (2014) operationalized it. The developer role broadened to absorb deployment, monitoring, and incident response — the "full-stack" era.

    Work toolChanging equipment
  • GitHub Copilot — first AI pair-programmer

    GitHub announced Copilot (technical preview) on June 29, 2021, built on OpenAI Codex. The first commercial product framed explicitly as "an AI that writes code with you." General availability followed June 21, 2022 at $10/month. GitHub later reported Copilot was "writing nearly 50% of code in files where it's enabled" — the first widely-cited substitution claim that wasn't hype.

    AI audit toolsPattern detection
  • ChatGPT — general LLM code assistance

    ChatGPT launched November 30, 2022. Within weeks developers were routing Stack Overflow questions through it instead of Stack Overflow; by 2024 the Stack Overflow Developer Survey reported 63.2% of professional developers using AI tools in their workflow (up from 44% in 2023, with 82% using AI for code generation specifically).

    Effect on the work

    Sentiment turned mixed fast — favorability dropped from 77% (2023) to 72% (2024); 45% of developers say AI handles complex tasks poorly, 31.2% explicitly distrust AI output.

    AI audit toolsPattern detection
  • Cursor + AI-first IDEs

    Cursor (Anysphere, 2023) was the first IDE rebuilt around an LLM rather than retrofitted with one. Within a year it had displaced VS Code + Copilot as the modal "AI coding workflow" for ambitious developers — the editor where the AI is the primary interface, not a sidebar.

    Work toolChanging equipment
  • Autonomous coding agents arrive (Devin)

    Devin (Cognition Labs, announced March 12, 2024) was the first product pitched as a fully autonomous software engineer: take a ticket, write code, run tests, open a PR, with no human in the loop. Cognition raised at a $350M valuation within weeks and turned down larger offers. Independent reviewers (Internet of Bugs) showed Devin failing on real Upwork tasks, and the autonomous-engineer pitch ran well ahead of what the tools could actually deliver. But the direction of travel was set, and within a year every major vendor shipped an "agent mode."

    Effect on the work

    The 2024 story was substitution: a machine pitched to do the whole job. In practice the early agents still needed close supervision, so the immediate effect was hype and anxiety far more than displacement.

    Work toolChanging equipment
  • Agentic coding goes multi-vendor: Claude Code, OpenAI Codex, Cursor, Copilot

    By 2025 and 2026 the daily reality was not one autonomous engineer but a crowded, multi-vendor field of agentic assistants a developer drives. The JetBrains AI Pulse survey (January 2026, over 10,000 professional developers) found roughly 90% regularly using at least one AI tool at work, led by GitHub Copilot (29%), with Cursor and Claude Code tied at 18% (Claude Code reached 24% in the US and Canada and posted the highest loyalty on the market at 91% satisfaction). OpenAI Codex was a striking case of how fast position can change. It launched early, an open-source terminal agent in April 2025 and a cloud agent in May 2025 on the codex-1 model, yet as late as January 2026 it sat at just 3% workplace usage and 27% awareness. Then it surged: GPT-5-Codex (September 2025) could run unattended for hours on a single task, GPT-5.3-Codex and a Cerebras-hosted "Spark" variant (February 2026) cut latency about 15x, a desktop app shipped to supervise several agents at once, and Codex crossed 2 million weekly active users by mid-March 2026, roughly fivefold growth in a quarter. The incumbents moved too: Cognition, the maker of Devin, bought the Windsurf IDE (closed December 2025) and shipped its own SWE-1.5 model, consolidating rather than replacing engineers, and Google entered with the Antigravity editor (November 2025, 6% adoption by January 2026). The common thread: the editor or terminal becomes a place you delegate work and review it, not type it line by line.

    Effect on the work

    Usage rose while trust fell. The Stack Overflow 2025 Developer Survey found 84% of respondents use or plan to use AI and about half of professional developers use it daily, yet favorability slid to 60% (61% among professionals), down from over 70% in 2023 and 2024. Only about a third trust the accuracy of AI output (33%) against 46% who actively distrust it, and the most-cited frustration was "AI solutions that are almost right, but not quite" (45%), with 66% saying they now spend more time fixing that almost-right code. Underneath the sentiment the substitution is measurable: by late 2025 the Anthropic Economic Index found that "modifying software to correct errors," debugging, was the single most common task in Claude API transcripts, and Computer and Mathematical tasks made up 46% of first-party API traffic. The work is shifting from writing code to specifying and reviewing it.

    AI audit toolsPattern detection
Projection cone · present → 2034

What credible sources project

Scrub the slider past now to anchor each scenario on the scrubber. The spread is the range of futures credible sources project for this role.

Employment outlook
Projected change in the number of people doing this work.
McKinsey Global Institute (2023)
2030
+8%
McKinsey's July 2023 "Generative AI and the Future of Work in America" report places STEM occupations (which includes software developers) as one of the few categories where the midpoint scenario is *positive* net employment growth through 2030, even after accounting for AI substitution — typically +5% to +15% by 2030 depending on adoption pace. Productivity gains in software engineering specifically are cited at 25-30% from generative AI tooling; the net effect on demand is positive at the modeled adoption rates.
BLS Occupational Outlook 2024
2034
+7%
BLS Employment Projections — industry-occupation matrix + labor productivity assumptions. The current published outlook for 15-1252 (2024-34 cycle): +7% employment growth ("Much faster than average"), 115,200 projected openings. Note: the prior cycle (BLS OOH 2023-33) projected +17% — the most recent revision is a sharp downward step that almost certainly reflects the post-2023 AI shock visible in the Computer Programmers (15-1251) series.
AI task exposure
Share of the role’s tasks that researchers estimate AI can do. This is a measure of task exposure, not a forecast of jobs lost.
Eloundou et al. — "GPTs are GPTs" (2023, observational)
2024
47%
of tasks
GPT-4 task-by-task labeling against O*NET task statements for 15-1252. Software developers are among the highest-exposure occupations in the set — conventional citation of β (E1 + 0.5×E2) ≈ 0.47, γ (any exposure) ≈ 0.78. Reported here as β = -47% to fit the cone display; γ would read as -78%. Note that "exposure" is *capability* not *substitution* — it is the share of tasks an LLM could do, not the share of jobs that will go away. Treat as ceiling.
Anthropic Economic Index (live observational)
2026
46%
of tasks
Direct measurement of Claude API usage by task category, Jan 2026 report. Computer & Mathematical tasks = 46% of all 1P API traffic in Nov 2025, up from 44% in Aug 2025. "Modifying software to correct errors" — debugging — is 10% of all API records, the single most common task. Reported here as -46% on the cone to represent *current usage share*, not a forecast; it directly maps to "AI is doing this fraction of the work in production today" rather than a future scenario.
Goldman Sachs (March 2023)
2030
29%
of tasks
Goldman maps O*NET work-activity importance scores to LLM capability ratings. Their March 2023 "Potentially Large Effects of AI on Economic Growth" report identifies Computer & Mathematical occupations as having ~29% of tasks automatable by current LLM capabilities. As with Eloundou, this is share of *tasks automatable*, not jobs lost — interpret as ceiling, not floor.
Frey & Osborne (2013)
2033
4%
of tasks
Gaussian-process classifier on O*NET task features. F&O classified the closest-comparable pre-2018 SOC (15-1132 Software Developers, Applications) as one of the LEAST computerisable jobs in the entire 702-occupation set — conventional citation of the appendix value is ~0.042 probability of computerisation, ~24:1 against. The sibling 15-1133 (Systems Software) was higher at ~0.13, and 15-1131 (Computer Programmers) was 0.48 — the divergence has held up. Exact appendix decimals could not be re-extracted in this curation pass (Frey & Osborne PDF table layout); the 4% figure is the lower-bound Applications value.
Today, in this role

What's shifting in the work right now

The historical view above shows how this role has moved. This is the present-day detail: which AI tools are picking up which tasks, where the edge still is, and the natural directions this work can grow.

What's changing in your day

Three parts of your work where AI is already doing real lifting, and what stays yours.

AI is taking this onMigrate and modernize legacy codebases — refactoring deprecated patterns, upgrading dependencies, and translating between languages or frameworks — tasks where autonomous agents (Devin, Copilot Workspace) now handle most repetitive transformation work.

Migrate and modernize legacy codebases — refactoring deprecated patterns, upgrading dependencies, and translating between languages or frameworks — tasks where autonomous agents (Devin, Copilot Workspace) now handle most repetitive transformation work.[7],[10]

Where your edge is

Redirect effort from mechanical refactoring to post-migration validation: define the correctness criteria, own the test harness, and perform the business-logic review that verifies the automated agent preserved intended behavior.

AI is sitting alongside you herePrototype new features or proof-of-concepts rapidly using AI-native editors and agent tools, iterating on natural-language specifications before committing to a full implementation plan.

Prototype new features or proof-of-concepts rapidly using AI-native editors and agent tools, iterating on natural-language specifications before committing to a full implementation plan.[11],[12]

Where your edge is

Become fluent in prompt engineering for code — writing precise, scoped natural-language specs that produce useful first drafts; treat AI-generated prototypes as spike artifacts requiring design review before production use.

AI is sitting alongside you hereWrite and maintain automated test suites (unit, integration, end-to-end), delegating test-case generation to AI tools while curating edge cases that reflect real production failure modes.

Write and maintain automated test suites (unit, integration, end-to-end), delegating test-case generation to AI tools while curating edge cases that reflect real production failure modes.[6],[3]

Where your edge is

Focus on test strategy rather than test authorship: design the test pyramid, define coverage standards, and own the decision of which failure modes matter most — AI can generate hundreds of tests but needs human guidance on what to cover.

Where this role is heading

Natural next steps for someone with your foundation: not exits, evolutions.

A direction you could grow

Computer and Information Systems Managers

The 2025-2026 labor market shows rising demand for managers who can evaluate AI tooling, set adoption strategy for agentic coding agents, and lead teams navigating AI-assisted development. Experienced developers who build management skills command the highest compensation premiums in the sector.

What you'd add
· AI adoption strategy: evaluating agentic tools (Devin, Copilot Workspace) for team ROI
· Budget ownership: headcount planning, cloud cost management, tool licensing
· Executive communication: translating technical risk into business impact
What it takesA real upskill, but a natural one
Share this year
Drops anyone you send it to straight into 2026.
Preview card
Part of Tech & Data · see all 28roles →
Different role?

See the same long-arc view for your own profession.

Browse the directory by industry, or search by title or SOC code. New roles ship every few weeks. Every profile cites every claim.

Browse all roles

The data behind this timeline

On record since1945
Latest tracked employment1,687,890 (US, 2025)
Latest median pay$135,980 (2025)
Outlook+7% by 2034 (BLS Occupational Outlook 2024)
View all 14 cited data points
YearUS employmentMedian annual paySource
19456n/aESTIMATE
195715,000n/aESTIMATE
1980318,000n/aESTIMATE
1990n/a$34,000BLS-CPS
1999n/a$64,031CENSUS
2000707,000n/aESTIMATE
2008n/a$85,430BLS-OEWS
20181,148,560n/aBLS-OEWS
2019n/a$105,590BLS-OEWS
20211,364,180$120,730BLS-OEWS
20221,534,790$127,260BLS-OEWS
20231,656,880$132,270BLS-OEWS
20241,693,800$133,080BLS-OEWS
20251,687,890$135,980BLS-OEWS
Embed this timeline on your site

Free for any site. Paste this where the timeline should appear; it stays interactive, every datapoint stays cited, and it sets no cookies on your page. How embedding works

<iframe src="https://futurehistory.earth/embed/15-1252"
  width="100%" height="430" style="border:0"
  title="Software Developers, a Future History timeline"
  loading="lazy"></iframe>

See all roles in Tech & Data