Coding Assistants and IDEs: from Visual Studio to AI agents

Software was written in heavyweight desktop IDEs installed from a disc or a license server. Microsoft shipped Visual Studio 97 as a bundle of its compilers and editors, and IBM's Eclipse went open source in November 2001. On Unix boxes, vi and Emacs did the same job with less chrome and more muscle memory.

The smartest feature of the era was IntelliSense, which Microsoft introduced in 1996 in a Visual Basic 5.0 preview. It completed member names and showed parameter lists from the type system. It knew syntax and symbols. It knew nothing about intent.

Everything else was manual. Engineers searched printed manuals, MSDN discs and mailing lists, kept code in CVS or Subversion on a company server, and reviewed changes over email or in a meeting room. Builds ran on someone's machine, and "works on my machine" was a real diagnosis.

What broke was knowledge transfer. A new hire needed weeks to learn a codebase, and the only index was the senior engineer who wrote it.

Two launches in 2008 reshaped the work. GitHub turned version control into a social, hosted product built around the pull request, and Stack Overflow, which opened in September 2008, became the place every engineer searched before thinking. Copying a vetted answer was the era's version of code generation.

The editor got lighter and free. Microsoft announced Visual Studio Code in April 2015, and by the 2025 Stack Overflow survey about 76% of respondents used it. JetBrains built the paid power-user alternative. Licenses gave way to subscriptions and per-seat plans, the shift I wrote about in the obituary for owning software.

Microsoft bought GitHub in 2018, and CI services ran the tests on every push. Cloud IDEs such as Codespaces put the whole environment in a browser tab.

What broke was attention. Engineers spent their day jumping between the editor, Stack Overflow, docs, tickets and review queues. The tools were better connected, but a human still wrote every line.

GitHub announced Copilot as a technical preview in June 2021 and made it generally available in June 2022 at $10 a month. It ran on OpenAI Codex, a GPT-3 variant trained on code, and predicted the next lines from the file you had open. For the first time the editor wrote code you had not typed.

ChatGPT arrived in November 2022 and moved the Stack Overflow habit into a chat window. Engineers pasted errors, asked for a function, and pasted the answer back. The shift shows up in the numbers: questions submitted to Stack Overflow fell 78% between December 2024 and December 2025.

The human still drove. Assistants saw one file or one chat, had no way to run the code, and were confidently wrong often enough that every suggestion needed reading. Cursor, a VS Code fork, pushed the editor toward whole-repo context and multi-file edits.

What broke was trust. Code got written faster, but the review burden moved onto the person who accepted the suggestion, and hallucinated APIs and packages slipped into real codebases.

In 2025 the assistant became an agent. Anthropic previewed Claude Code in February and made it generally available in May. OpenAI shipped the open-source Codex CLI in April and a cloud Codex agent in May. GitHub's Copilot coding agent, announced at Build, reached general availability in September 2025: you assign it an issue, it works in a GitHub Actions environment, and it opens a draft pull request for review.

In production today, agents read the repo, plan, edit many files, run the tests, fix what fails, and hand back a PR. GitHub lets teams assign the same issue to Copilot, Claude or Codex. Cursor, Devin and Kiro run fleets of cloud agents in parallel sandboxes for hours. Cursor's own site says its agents "build, test, and demo features end to end for you to review." My practical notes on what actually works with AI in software development in 2026 cover where this pays off.

The money followed. Anthropic said in February 2026 that Claude Code had passed $2.5 billion in run-rate revenue. SpaceX closed its $60 billion acquisition of Cursor in August 2026, and Cognition, which bought Windsurf in July 2025 after Google hired its founders, passed $1 billion in run rate in September 2026.

The limits are real. METR's 2025 trial found experienced open-source developers were 19% slower with early-2025 tools while believing they were faster, and Stack Overflow's 2025 survey found 46% of developers distrust AI accuracy. Agents still need a human to approve the merge, and the practitioner playbook for QA and security teams is mostly about building the guardrails around them.

My bet is that by 2031 most routine code changes in well-run teams start as a written spec or an issue and finish as an agent-authored pull request. Engineers will spend their time deciding what to build, writing acceptance criteria, and reviewing. That is the shift the Future Tech entry on coding in English predicts, and the 2026 market already points that way.

The interface moves from the editor to the queue. Devin Desktop, Cursor and GitHub's agent views already treat the IDE as a control room for many agents. Expect review to be agent-assisted too, with a second model checking the first and a human signing off on what matters.

At the edges, software becomes disposable. Small internal tools will be generated for one task and thrown away, the idea behind just-in-time software.

For this to happen, three things must get better: verification that scales with output, agent identity and permissions that security teams can audit, and pricing that is cheaper than an engineer's hour. If verification stalls, the bottleneck simply moves to review and the gains flatten.