Every year someone publishes a list of software development trends, and every year most of the list is wrong by December. This one is different because 2026 has actual survey numbers behind it, from JetBrains, Stack Overflow, BairesDev, Gartner, Devographics, and DX. Four of them are large enough to draw conclusions from.
Coding agents went from experiment to default in roughly eighteen months. How teams review that code, and what they pay per developer for it, is still being worked out.
Adoption is no longer the interesting number
JetBrains surveyed more than 15,000 professional developers between April and May 2026. Ninety percent use AI coding agents at work at least weekly. Sixty-eight percent use them daily. Those numbers settle the question of whether agents belong in a professional workflow. They do.
The vendor landscape shifted just as fast. Claude Code went from 18% adoption in January 2026 to 39% by May, and 47% in the United States. It is now used roughly twice as often as GitHub Copilot, whose share actually fell from 29% to 21% over the same period. Codex climbed from zero to 16%, passing Cursor at 12%. JetBrains' own Junie slipped from 13% to 9%.
Stack Overflow's April 2026 pulse survey, about 1,100 respondents, found agentic use at work had nearly doubled since their 2025 survey, from 31% to 59%.
Most teams still keep the agent on a short leash
Enthusiasm and behaviour diverge sharply here. In the same Stack Overflow pulse, 68% of respondents preferred a single agent over multi-agent orchestration, and 63% said they rarely or never let an agent run fully on autopilot.
The single-agent preference has a technical reason. Single-agent setups keep the call stack shallow, which shows up as lower latency. Stack Overflow measured an average of 850ms against 1.4 seconds for multi-agent chains. One token scope to audit instead of many, and fewer hand-offs where a prompt injection can take hold.
The multi-agent users, roughly 32% of respondents, are power users who orchestrate specialised agents for retrieval, generation, and testing. They report much higher daily usage at 70%, along with higher latency and a wider security surface.
How much of your code is actually agent-written
This is the number vendors and practitioners both tend to get wrong.
JetBrains asked. The median response was about 10% agent-generated, 25% AI-assisted, and 75% written by hand. BairesDev's survey found the share of developers using AI to write half or more of their code jumped from 12% to 42% year over year, which is a much more aggressive picture.
Both can be true. A small number of developers have restructured their entire workflow around agents while the median professional still writes most code by hand, reaching for AI when stuck.
Trust is moving the wrong way
Stack Overflow's 2025 survey of nearly 50,000 developers found 84% use or plan to use AI tools, up from 76% in 2024. Over the same period, the share who trust AI output to be accurate fell from 40% to 29%. More developers actively distrust the accuracy (46%) than trust it (33%), and only 3% report high trust.
ProLLM research explains why. On unseen real questions pulled from Stack Overflow, leading models scored under 50% correct, with GPT-4o at 45.5% and Claude Sonnet 3.5 at 47.5%. Models agreed with incorrect outputs up to 72.5% of the time.
One engineer put it well on Stack Overflow's podcast: every error you hit erodes your trust a little further. The result is that developers lean on people when they do not trust the answer, with 75% turning to another developer, and over 80% still visiting Stack Overflow regularly despite the tooling.
The other signal: advanced technical questions on Stack Overflow have doubled since 2023. Hard problems are exactly where these tools fall short.
The human checkpoint is where the money went
BairesDev asked 41 CTOs. Seventy-eight percent had increased spending on code review, quality assurance, and validation to support AI-generated work. Only 7% of developers said the decision to ship code had been entirely delegated to AI without their input.
Generation got cheap, so verification became the bottleneck. An Amazon principal engineer told Stack Overflow that agentic engineering is pushing code bottlenecks downstream into testing and deployment, and that robust validation is what makes fearless commits possible.
Test coverage and review capacity are where the return is, ahead of another seat.
Token costs are the new line item
The shift from per-seat licences to consumption pricing caught a lot of teams off guard. Gartner Peer Insights found 23% of tech leaders spending $200 to $500 per developer per month on tokens. Six percent pay more than $2,000. Gartner forecasts that by 2028, AI costs will overtake the average developer salary.
DX puts the total cost, seat plus tokens, at $200 to $600 per developer per month for teams mixing inline and agentic tools.
The anecdotes are more alarming than the averages. TechCrunch reported in June 2026 that Uber blew through its entire 2026 AI coding budget by April, Microsoft revoked developer Claude Code licences months after enabling them, and one company found itself with a $500 million bill after forgetting to set usage limits. A CTO told Faros AI that one of his engineers spent $40,000 on tokens in a month, and he did not know whether to stop him or tell everyone else to do the same.
Productivity data is murkier than the spending implies. Jellyfish found engineers using the most tokens were about twice as productive as lighter users, but spent 10 times the tokens to get there, with per-developer consumption rising roughly 18.6x in nine months. Faros' two-year study of 20,000 developers found output rising alongside bugs and rewrites. DX, tracking more than 400 organisations over 14 months, measured a median PR throughput gain of 7.76% and a mean of 13.1%.
A 7.76% median gain is worth having, and it is a long way from the order of magnitude being advertised.
What is emerging in response is FinOps for tokens. The Linux Foundation set up a Tokenomics Foundation in June 2026 to apply cloud cost discipline to AI token spend.
TypeScript won the language argument
The State of JavaScript 2025 survey collected 13,002 responses between September and November 2025. Forty percent of respondents now write only TypeScript, up from 34% in 2024 and 28% in 2022. Only 6% write plain JavaScript exclusively. Despite that, missing static typing was still the number one language pain point.
Runtime type stripping in stable Node.js versions means TypeScript annotations can ship without a compile step, which Daniel Roe of the Nuxt core team described as TypeScript winning as a language rather than as a bundler.
Build tools turned on Webpack
Webpack still leads on usage at 86.4% against Vite at 84.4%, a two-point gap. Sentiment tells a different story. Vite has 56% positive sentiment and 1% negative. Webpack has 14% positive and 37% negative. That is a 78-point satisfaction differential, and Vite now has Webpack's usage with none of the resentment.
Rolldown, the Rust-based Rollup replacement, jumped from 1% to 10% adoption in one year. Stable Rolldown is expected to power a faster Vite in 2026.
Front-end frameworks stopped churning
React remains the most used framework at 83.6%. Next.js sits at 59% usage with 21% positive and 17% negative sentiment, which generated more commentary than any other project in the survey, including complaints about mounting complexity. Solid.js has held the highest satisfaction rating for five consecutive years, and Astro continues to gain in the meta-framework space.
Overall happiness held at 3.8 out of 5 for a fifth straight year. The ecosystem is settling rather than churning. Svelte, the survey organisers noted pointedly, is nine years old.
What to do with this
Fund verification before you add tooling. If your team is generating more code, the binding constraint is review and test capacity. Measure throughput before and after adoption so you know whether the spend is justified.
Set token budgets before you scale. The organisations that got a $500 million surprise had no limits configured. Track cost per developer per month the way you would track cloud spend.
Keep one agent on routine work. Multi-agent orchestration costs latency and widens the security surface, so reserve it for work where the parallelism pays for itself.
Adopt TypeScript. The move from 28% of developers in 2022 to 40% in 2025 is the clearest consensus signal in the ecosystem, and runtime type stripping removes the last real objection.
Plan on a modest gain. Most credible measurements land between 5% and 15% throughput improvement, with DX's 400-company study putting the median at 7.76%. If someone promises you ten times that, ask which metric they measured and over what period.