How AI agents change developer jobs: 2025, 2026 and a 2027 guess

Will AI agents replace software developers? The short answer

Three lines, then the numbers.

AI now writes most of the new code in teams that use it: Google says three quarters of its new code (Google, April 2026), Anthropic says more than 80% of what it merges (Anthropic, June 2026). Job ads for developers in the US are still well below 2020 (Indeed, via FRED), and jobs for people who build with AI are the fastest growing on LinkedIn (LinkedIn, January 2026). The job is changing from writing code to directing and checking agents. Nobody’s numbers say it disappears.

Here are the numbers, year by year, both directions. I build with agents every day, so I read them as a practitioner.

2025: the year AI started writing most of the code

84%of developers used or planned to use AI toolsStack Overflow, July 2025

66%said the top frustration was answers that are “almost right, but not quite”Stack Overflow, July 2025

No. 1TypeScript became the language with the most monthly contributors on GitHubGitHub Octoverse, October 2025

13%lower employment for 22 to 25 year olds in the most AI-exposed jobs, by July 2025Stanford, via LeadDev, August 2025

In 2025 the assistant sat inside the editor and suggested lines. Stack Overflow’s survey of that summer found 84% of developers using or planning to use AI tools, 51% of professionals using them daily (Stack Overflow, July 2025), and 31% already using agents (Stack Overflow blog, September 2026). Trust was low: 33% trusted the output, 46% did not, and the top frustration, named by 66%, was answers that are “almost right, but not quite”.

The companies went further than the surveys. Google said in October 2025 that “nearly half” of its new code was AI-generated (Google, October 2025), up from “more than a quarter” a year earlier (Fortune, October 2024). On GitHub, TypeScript overtook Python and JavaScript in August 2025 as the language with the most monthly contributors; GitHub’s own reason was that typed languages make “agent-assisted coding more reliable in production” (GitHub Octoverse, October 2025).

The job market told two stories at once. US software job postings on Indeed hit their lowest point in May 2025, about 40% below the level of February 2020 (Indeed, via FRED). Stanford’s payroll study found employment of 22 to 25 year olds in the most AI-exposed jobs, software development among them, down 13% by July 2025 while overall US employment kept growing (LeadDev on the Stanford study, August 2025). In the same year, Amazon said AI would “reduce our total corporate workforce” in the coming years (Amazon, June 2025), and LinkedIn’s data on job moves up to mid 2025 put AI engineer first among the fastest-growing jobs (LinkedIn, January 2026).

2026: the year of agents

73%of developers who use AI coding tools use them dailyStack Overflow, October 2026

80%+of the code Anthropic merges is written by ClaudeAnthropic, May 2026 figure

~5 hoursthe length of task the best agent finished half the timeMETR, January 2026

4.5%of US software job postings are entry-levelIndeed Hiring Lab, Q1 2026

In 2026 the assistant became a worker. Stack Overflow’s new survey, out this month, says 73% of developers who use AI coding tools use them daily, 80% spend at least an hour a day with AI, and about 70% hand tasks to an agent (Stack Overflow, October 2026). JetBrains measured 90% of developers using coding agents at least weekly by mid 2026 (JetBrains, August 2026).

The output numbers jumped. Google reported 75% of its new code AI-generated in April 2026 (Google, April 2026). Anthropic reported that more than 80% of the code merged into its own codebase in May 2026 was written by Claude, and that a typical engineer merged eight times as much code per day as in 2024; the same report warns that lines of code measure quantity, not quality (Anthropic, June 2026). Across ordinary companies the share is lower: a Sonar survey of 1,100 developers put AI-written or assisted code at 42% in January (Sonar, January 2026), and Veracode’s 2026 report says “roughly half” (Veracode, July 2026). METR, which measures how long a task an agent can finish half the time, put the best model at about five hours in January 2026, with the length doubling about every four months (METR, January 2026).

The job market, both directions, in one paragraph. Down: US software postings in October 2026 are still about a fifth below February 2020 (Indeed, via FRED); employment in US computer systems design firms is about 5% below its March 2023 peak and has fallen slowly every half year since (BLS, via FRED); Stanford’s revised study of June 2026 payrolls puts young workers in AI-exposed jobs 19% below where they would be if they had kept pace with their peers, mostly through less hiring, not firing (Stanford Digital Economy Lab, August 2026); Indeed finds entry-level roles are now 4.5% of software postings, the lowest share of any sector it tracks (Indeed Hiring Lab, July 2026); and tech layoffs continued, Amazon’s the largest (LeadDev, February 2026). Up: Indeed also reports US software postings rose almost 15% in the year after Claude Code launched, with 71% of the rise in senior roles and 37% in jobs with AI in the title (Indeed Hiring Lab, July 2026); many large tech companies list 50 to 100% more AI engineering jobs than a year ago (The Pragmatic Engineer, May 2026); and AI engineer is the number one rising job on LinkedIn’s 2026 list, in the US and in Australia (ACS Information Age on LinkedIn’s list, January 2026).

My reading, marked as a reading: the number of developer jobs is not falling off a cliff. It is tilting towards senior people and towards people who direct agents, and the entry door is narrower.

2027: my guess, and what it rests on

65%of committed code written by AI, what developers expect by 2027Sonar survey, January 2026

80%of the engineering workforce will need to upskill through 2027Gartner, October 2024

~4 monthsfor the length of tasks agents can finish to doubleMETR, January 2026

+10%US software developer jobs, projected 2025 to 2035US Bureau of Labor Statistics, August 2026

This is a guess. Here is what it stands on.

Sonar’s developers expect AI to write 65% of committed code by 2027 (Sonar, January 2026). Gartner said in 2024 that 80% of the engineering workforce would need to upskill by 2027 and that the agent phase leads to “AI-native software engineering” where most code is generated rather than typed (Gartner, via CDO Magazine, October 2024). METR’s doubling trend, if it holds, means agents that finish day-long tasks half the time within two years (METR, January 2026). The US Bureau of Labor Statistics still projects software developer employment to grow 10% from 2025 to 2035, with about 106,100 openings a year (BLS, August 2026).

So my 2027 Tuesday looks like this. The developer plans in the morning, an agent team does most of the typing, the developer reads diffs and test results, and one person approves the deploy and anything that touches money. Teams ship more with the same people. The entry door narrows further unless teams open it on purpose by pairing juniors with agents. The people hired first are the ones who can brief, review and direct. If the numbers above move, this guess moves with them.

How much code AI writes: 2025, 2026, 2027

Share of code written by AI, by sourceEach source measures something different, so each has its own rows. Compare the direction, not the size.

• Earlier years
• 2026
• Expected, not measured

See the numbers

What	Number	Source
Google, new code generated by AI, 2024	more than a quarter	Fortune, 30 Oct 2024
Google, new code generated by AI, 2025	nearly half	Google, 9 Oct 2025
Google, new code generated by AI, 2026	75%	Google, 22 Apr 2026
Anthropic, merged code written by Claude, before February 2025	low single digits	Anthropic, June 2026
Anthropic, merged code written by Claude, May 2026	more than 80%	Anthropic, June 2026
Sonar survey, committed code from AI, 2026	42%	Sonar, 8 Jan 2026
Sonar survey, committed code from AI, expected by 2027	65%	Sonar, 8 Jan 2026
Veracode, committed code from AI, 2026	roughly half	Veracode, 28 Jul 2026

Different sources measure different things, so compare direction, not size. Company claims: Google a quarter (2024), nearly half (2025), three quarters (2026); Anthropic low single digits before February 2025, more than 80% by May 2026. Surveys of ordinary teams: 42% (Sonar, January 2026) and roughly half (Veracode, July 2026). The 2027 bar is Sonar’s 65%, which is what developers expect, not a measurement.

Developer job ads 2025 to 2026: what is down and what is up

DownFewer ads than before 2020, and fewer first jobs.

~1/5US software postings below February 2020Indeed via FRED, October 2026

4.5%of US software postings are entry-levelIndeed Hiring Lab, Q1 2026

19%young workers in AI-exposed jobs behind their peersStanford, August 2026

~5%US computer systems design jobs below the March 2023 peakBLS via FRED, September 2026

UpMore ads than a year ago, mostly senior and AI roles.

~15%rise in US software postings, May 2025 to May 2026Indeed Hiring Lab, July 2026

No. 1AI engineer, the fastest-rising job in the US and in AustraliaLinkedIn, January 2026

50–100%more AI engineering listings at many large tech companiesThe Pragmatic Engineer, May 2026

+10%US software developer jobs, projected 2025 to 2035BLS, August 2026

Down: US software postings on Indeed are below their 2020 level in both years, with the low point in May 2025 and a slow climb since (Indeed, via FRED); entry-level is 4.5% of software postings; young workers in AI-exposed jobs are 19% behind their peers. Up: software postings rose almost 15% in the year to May 2026; AI engineer is the number one rising job; the long US projection is still plus 10% over ten years. Both lists are true at once. The work moved, and it moved up the ladder.

−0.3%Australia’s tech workforce in 2025, the first fall on recordACS Digital Pulse, August 2026

216,000software and applications programmers at work, a recordABS data via Adaca, June 2026

+10.1%ICT job ads year on year; 13.4% of them mention AISEEK, February 2026 data

Australia is its own picture. The tech workforce fell for the first time on record in 2025, by 0.3% to about 967,000, with the losses in sales, support and administration roles and gains in technical and professional roles (ACS, August 2026). At the same time the number of employed software programmers reached a record 216,000 in 2026, up from 189,000 (IT Brief on ABS data, June 2026). SEEK’s ICT job ads were up 10.1% year on year in February 2026, and 13.4% of them mention AI (SEEK, March 2026). Indeed’s Australian software index sits a little above its 2020 level but below where it started the year (Indeed, via FRED). ACS says the country needs 259,000 more technology workers over the next decade.

Programming languages in 2026: which ones AI agents write most

Languages developers use, 2025 and 2026Stack Overflow Developer Survey, all respondents. Share who used each language in the past year.

• 2025
• 2026

Nobody publishes which languages the agents write most; no vendor has released that data. What we can see is what developers use. Stack Overflow 2026: Python 58%, TypeScript 44%, Rust 18% (up from 15%), Go 16.5% (Stack Overflow, October 2026). GitHub’s contributor count put TypeScript first in August 2025 and tied the rise to agents (GitHub Octoverse, October 2025). My reading: agents write the language you can read and test fastest, and typed languages make their mistakes visible sooner.

What to do now: a developer, and a team lead

If you write code

• Learn to brief an agentA clear task, the files it may touch, how you will check it.
• Review the diff, read the testsThat is the job the ads describe now.
• Keep one deep skillThe one the agents cannot fake.
• Ship with agents every dayNot once a month.

If you lead a team

• Measure before and afterSpeed and stability, not only speed.
• Put tests and review in front of the agentsNot behind them.
• Pair juniors with agentsOpen the junior door on purpose.
• Name one owner for the agent setupOne person, clearly named.

If you write code: learn to brief an agent, review its diff and read its tests, because that is the job the ads describe. Keep one deep skill the agents cannot fake. Ship with agents every day, not once a month.

If you lead a team: measure before and after, because DORA found AI raises throughput and lowers stability in the same teams (DORA, September 2025). Put tests and review in front of the agents, not behind them. Open the junior door on purpose by pairing juniors with agents, or the 4.5% becomes your team’s number too. Name one person to own the agent setup. That person is the role I take when I join a team.