Wave 1: ML Completion (2018–2021)
IDE-integrated autocomplete, bundled into existing tools
Wave 2: LLM Coding (2021–2024)
Copilot, ChatGPT, and chat-based code generation
Wave 3: Agentic Systems (2024–)
Autonomous agents that plan, execute, and submit PRs
2018 2 events
Visual Studio IntelliCode announced +

Microsoft brings AI-assisted IntelliSense to Visual Studio — the first mainstream IDE to ship ML-powered code completion.

IntelliCode used machine learning models trained on thousands of open-source projects to provide contextual code suggestions. It was bundled into Visual Studio, making "AI-assisted development" accessible without extra installation — but most developers didn't think of it as "AI."
Tabnine launches first AI code assistant +

The first dedicated AI code completion tool ships as a standalone IDE extension.

Originally called "TabNine," it was one of the first products to apply deep learning specifically to code completion. Vendor later claims "more than one million users and thousands of organizations" — though not independently audited.
2021 2 events
GitHub Copilot technical preview +

The catalyst — LLM-powered code generation enters the IDE for the first time at scale, powered by OpenAI Codex.

Built on OpenAI's Codex model (a fine-tuned GPT-3), Copilot could generate entire functions from comments and context. The technical preview was invitation-only but generated enormous developer interest and debate about AI-generated code quality and copyright.
OpenAI Codex research paper published +

The model behind Copilot is described in academic literature (arXiv:2107.03374).

Codex demonstrated that LLMs could solve programming problems at a meaningful rate, establishing code generation as a legitimate capability of large language models and setting the stage for the next wave of products.
2022 2 events
GitHub Copilot generally available +

Copilot goes paid ($10/mo) — free for verified students and popular OSS maintainers, deliberately seeding the open-source community.

The free tier for OSS maintainers was a key go-to-market strategy, ensuring that influential open-source contributors would adopt and advocate for the tool. GitHub also published a controlled study: 95 developers using Copilot completed a task 55% faster on average (1h11m vs 2h41m).
ChatGPT introduced +

OpenAI's chat interface goes viral, broadening AI-assisted coding from IDEs into general-purpose conversation — a second adoption vector.

While not a coding tool per se, ChatGPT became the most widely used AI tool among developers (89% per HackerRank 2025). Its conversational interface made code generation, debugging, and explanation accessible to developers who hadn't adopted IDE extensions.
2023 5 events
ChatGPT plugins + Code Interpreter announced +

ChatGPT gains the ability to execute code in a sandbox — the first hint of agentic capability in a consumer product.

Code Interpreter could write and run Python, analyze data, create visualizations, and work with uploaded files. It rolled out broadly in July 2023 to Plus users, and demonstrated that LLMs could move beyond suggestion into execution.
Codex models deprecated in OpenAI API +

OpenAI retires the original Codex models — GPT-3.5/4 prove capable enough for code without a specialized model.

The deprecation signaled that general-purpose LLMs had subsumed code-specialized models. This complicated longitudinal analysis: "Codex" would later be reused as a brand name for OpenAI's 2025 agentic product, referring to a materially different system.
Stack Overflow: 43.8% AI tool adoption +

First major benchmark — 89,184 respondents; 25.5% plan to use soon. The baseline for all future comparisons.

This was the first large-scale survey to ask specifically about AI in the development process (narrower than "ever used"), establishing a measurable adoption rate. Combined use-or-plan rate was ~69%.
SWE-bench published (Princeton) +

The benchmark that would define the agentic coding era — 2,294 real GitHub issues from 12 Python repos.

SWE-bench (arXiv:2310.06770, Jimenez et al.) evaluates whether AI can resolve real-world software issues end-to-end. Accepted as an oral at ICLR 2024, it became the de facto standard for measuring agentic coding capability. By mid-2024, top agents scored ~20% on the full benchmark.
Sourcegraph Cody GA +

AI coding assistant leveraging Sourcegraph's deep code search for codebase-aware context launches generally.

Cody differentiated itself by using Sourcegraph's code intelligence graph for context. Cody Pro launched at $9/month. The product would later be pivoted — Cody Free/Pro sunset in July 2025 as Sourcegraph spun out "Amp" as a separate agentic coding company.
2024 10 events
Devin announced — "first AI software engineer" +

Cognition Labs demos an autonomous coding agent that can plan, execute, and debug multi-step tasks. Goes viral — and polarizes the industry.

Devin claimed SOTA on SWE-Bench at launch. Backed by Founders Fund. Industry reaction was split: enormous viral excitement at the demo, followed by skepticism when independent reviews showed underwhelming real-world performance. The Register later ran "'First AI software engineer' is bad at its job." GA came in December 2024 at $500/month; Devin 2.0 launched April 2025 at $20/month.
Google Gemini Code Assist launches +

Google rebrands Duet AI for Developers → Gemini Code Assist at Cloud Next 2024. Powered by Gemini 1.5 Pro.

Part of a broader February 2024 consolidation retiring all "Duet AI" branding in favor of "Gemini." Supports VS Code, JetBrains IDEs, and Cloud Shell Editor. An enterprise version with codebase-aware customization followed. Gemini later shows 47.4% usage among SO 2025 agent users.
Augment Code emerges from stealth — $252M raised +

Enterprise-focused AI coding startup reveals itself with a $977M valuation after two years of stealth development.

Founded in 2022, Augment focused specifically on AI coding for large enterprise codebases. Investors: Sutter Hill Ventures, Index Ventures, Lightspeed, Meritech Capital. Reached ~$20M revenue by October 2025.
Amazon Q Developer GA +

AWS rebrands CodeWhisperer → Amazon Q Developer, expanding from code completion to a full-lifecycle AI assistant.

All CodeWhisperer features were folded into Q Developer, which added broader capabilities: debugging, optimization, migration assistance, and infrastructure management. AWS claimed it could make employees "more than 80% more productive."
GitHub Copilot Workspace technical preview +

GitHub unveils an issue-to-PR workflow: AI that brainstorms, plans, implements, and validates code changes from a GitHub issue.

First teased at GitHub Universe 2023, Workspace entered technical preview on April 29, 2024. A "Copilot-native developer environment" where the AI proposes a full implementation plan and code diff from an issue description. Expanded to all paying Copilot customers in December 2024.
Stack Overflow 2024: adoption jumps to 61.8% +

65,437 respondents — AI tool use rose 18 points year-over-year. "No plans" drops from 29.4% to 24.4%.

Same survey framing as 2023, making this year-over-year comparison reliable. The combined "use or plan" rate reached ~76%. The holdout population shrank from ~30% to ~24% in just one year.
EU AI Act enters into force +

The world's most comprehensive AI regulation takes effect, beginning a phased enforcement timeline that reaches coding tools via GPAI model obligations.

The Act doesn't single out coding tools, but GPAI model obligations (August 2025) require providers to supply technical documentation and comply with EU copyright rules. Models posing "systemic risks" (>10²⁵ FLOPS) face additional adversarial testing requirements. Full applicability for most operators arrives August 2026.
Cursor raises $165M across Series A & B +

Anysphere (Cursor) raises $60M at $400M valuation (August), then $105M at $2.5B valuation (December) — a 6x jump in four months.

Led by a16z and Thrive Capital. Cursor reached ~$100M ARR in 2024. The back-to-back raises signaled explosive growth and positioned Cursor as the leading AI-native IDE challenger to VS Code + Copilot.
Replit Agent launches +

An autonomous agent that builds complete applications from natural language — not just suggestions, but full-stack app generation.

Replit Agent drove explosive growth: Replit went from $10M ARR (end of 2024) to $100M ARR (June 2025), with subscriber base growing 45% monthly post-Agent launch. Targeted a different user segment than Copilot/Cursor — non-professional developers and rapid prototyping.
Windsurf launches agentic IDE +

Codeium launches Windsurf Editor, branded as the "first agentic IDE" — combining copilot and agent paradigms in a single editor.

Windsurf combined inline completions with agent-style multi-file editing in one interface. By July 2025, Reuters reported $82M ARR and 350+ enterprise clients. The company fully rebranded from Codeium to Windsurf in April 2025.
2025 16 events
JetBrains Junie announced +

JetBrains enters the agent race with a coding agent built into their IDE ecosystem — routine task delegation from familiar tools.

Junie could handle routine coding tasks, tests, and refactoring within JetBrains IDEs. Market-wide penetration not publicly quantified, but adoption was likely concentrated within the JetBrains user base.
Claude Code research preview +

Anthropic introduces an agentic CLI tool that reads codebases, edits files, and runs commands — launched alongside Claude 3.7 Sonnet.

Claude Code took a fundamentally different approach: terminal-first rather than IDE-embedded. It could autonomously navigate repos, edit multiple files, run tests, and iterate. The launch was notably low-key — no launch event, no viral demo. It reached GA on May 22, 2025.
GitHub Copilot agent mode preview +

Copilot gains autonomous multi-step capability in VS Code Insiders — planning tasks, editing multiple files, running terminal commands, and self-correcting errors.

Agent mode was distinct from standard Copilot (inline completions) and Copilot Chat (Q&A). It could break down tasks, identify sub-problems, execute solutions across files, and iterate on errors autonomously. GA rolled out to all VS Code users ~April 2025. Extended to JetBrains, Eclipse, and Xcode by May 2025.
EU AI Act: prohibited practices + AI literacy enforceable +

First enforcement phase — organizations must comply with banned AI practices and demonstrate AI literacy among staff.

While coding tools themselves weren't in the "prohibited" category, the AI literacy requirement meant organizations using AI coding tools needed training and governance programs in place. This accelerated enterprise AI policy formalization.
Devin 2.0 — price drops from $500 to $20/month +

Cognition slashes pricing and shifts to usage-based model, signaling that autonomous coding agents are moving toward mass-market accessibility.

Devin 2.0 introduced ACUs (Agent Compute Units) for usage-based pricing. The 96% price drop in four months reflected both competitive pressure and the reality that early pricing had limited adoption to enterprises willing to experiment.
Tabnine sunsets free tier +

Tabnine discontinues its free "Basic" plan to focus entirely on enterprise — named a Visionary in 2025 Gartner Magic Quadrant for AI Code Assistants.

The free-to-enterprise pivot reflected a market where free-tier competition from Copilot, Gemini Code Assist, and others made individual freemium unsustainable. Tabnine doubled down on enterprise features: private deployment, IP protection, and compliance.
OpenAI Codex agent research preview +

OpenAI launches a cloud-based software engineering agent — a sandboxed system that can write features, fix bugs, and propose PRs autonomously.

Not to be confused with the deprecated 2021 Codex model. The 2025 "Codex" is a fully agentic platform that runs in sandboxed cloud environments, operates on entire codebases, and submits pull requests. The name reuse complicated longitudinal tracking.
GitHub Copilot coding agent public preview +

Announced at Microsoft Build — an autonomous agent that takes a GitHub issue, spins up a dev environment, writes code, and opens a draft PR.

Distinct from "agent mode" (interactive, in-IDE). The coding agent is asynchronous: assign it an issue and it works in the background via GitHub Actions. Excels at low-to-medium complexity tasks in well-tested codebases. Reached GA on September 25, 2025 for all paid Copilot subscribers.
Claude Code reaches general availability +

Three months after research preview, Claude Code goes GA. Becomes one of the fastest-growing software products ever: ~$1B ARR by November 2025.

Claude Code on the web launched October 20, 2025. Pragmatic Engineer's 2026 survey found 71% usage among regular agent users and 75% at the smallest companies — the dominant tool in the agentic-first segment. Estimated near $2B ARR by January 2026.
OpenAI–Windsurf acquisition collapses +

OpenAI agrees to buy Windsurf for $3B (May) — deal falls apart (July). Google signs $2.4B licensing deal and hires Windsurf's CEO.

The deal collapsed partly because Microsoft (via its OpenAI relationship) would have gained access to Windsurf's IP — problematic given Microsoft's competing Copilot product. Google DeepMind hired Windsurf CEO Varun Mohan and co-founder Douglas Chen. Windsurf continues independently with a non-exclusive Google license.
Stack Overflow 2025: 78.5% using, 46% distrust +

49,009 respondents — 47.1% daily use, but more developers distrust AI accuracy (46%) than trust it (33%). The paradox crystallizes.

2025 shifted to frequency reporting. Agent-specific data showed 30.9% using agents at any frequency, with 37.9% saying they don't plan to use agents. Only ~3% reported "highly trusting" AI output. Top frustration: "almost right but not quite" outputs (66%).
Cursor raises $900M at $9.9B — fastest-growing SaaS ever +

Series C co-led by Thrive and a16z. Cursor hit $500M ARR (May) → $1B ARR (October) — a pace unprecedented in SaaS history.

By November 2025, Cursor raised $2.3B Series D at $29.3B valuation. Total funding: ~$3.3B from Google, NVIDIA, a16z, Thrive, Accel, and Coatue. The valuation trajectory — $400M → $2.5B → $9.9B → $29.3B in 15 months — reflected the market's conviction that AI-native IDEs would replace traditional editors.
EU AI Act: GPAI model obligations take effect +

Providers of general-purpose AI models must supply technical documentation, publish training data summaries, and comply with EU copyright rules.

This milestone directly affected the foundation models powering all major coding tools (GPT-4, Claude, Gemini). Models posing "systemic risks" (>10²⁵ FLOPS) faced additional adversarial testing and risk mitigation obligations. The full compliance framework for most operators arrives August 2026.
Claude Code v2.0 — subagents, hooks, Agent SDK +

Major architecture upgrade: checkpoints, subagents, hooks, background tasks, and the Claude Agent SDK. Powered by Sonnet 4.5.

The v2.0 release marked the shift from Claude Code as a single-turn tool to an orchestration platform. The Agent SDK enabled developers to build custom agentic workflows, while hooks and checkpoints allowed programmatic control over long-running autonomous sessions.
Cursor 2.0 + multi-agent interface +

Cursor ships a multi-agent interface where multiple AI agents collaborate on different parts of a codebase simultaneously.

Cursor 2.0 represented the frontier of agentic IDE design — moving from single-agent to multi-agent workflows. This pushed the boundary on what "AI-assisted development" could look like: multiple agents working in parallel, each handling a different file or subsystem.
Devin: 67% PR merge rate, 4× faster problem solving +

Cognition Labs publishes Devin's 2025 performance review — PR merge rates doubled from 34% to 67%. Goldman Sachs pilots "hybrid workforce" with 20% efficiency gains.

Devin excels at tasks with clear requirements that would take a junior engineer 4-8 hours: migrations, vulnerability fixes, unit tests. One organization saved 5-10% of total developer time on security fixes alone. 20× efficiency gain on vulnerability remediation (1.5 min vs 30 min per issue).
2026 9 events
Claude Cowork announced — agents beyond coding +

Anthropic extends the agentic paradigm from coding to general knowledge work on macOS. Plugin system follows Jan 30.

Cowork signaled that the agentic infrastructure built for Claude Code was generalizable. The same long-horizon execution patterns — subagents, checkpoints, background tasks — could apply to research, analysis, and document creation.
METR: AI agent time horizons doubling every ~89 days +

METR publishes Time Horizon 1.1 — the "50% time horizon" (task length at which agents succeed half the time) jumped from ~4 min (early 2024) to hours by late 2025.

The Time Horizons benchmark tracked 228 tasks across frontier models. The doubling rate accelerated from ~188 days (full period) to ~89 days since 2024. This exponential improvement — described as "a new Moore's Law for AI agents" — was the strongest quantitative evidence that agent capabilities had crossed a practical threshold for long-horizon autonomous work.
Microsoft: 4.7M paid Copilot subscribers +

FY26 Q2 earnings call — the hardest number in AI coding. "20M users" and "90% of the Fortune 100" also cited.

Paid subscriber count is the most auditable metric in the space. The gap between "4.7M paid" and "20M users" suggests significant free/trial usage. Fortune 100 penetration (90%) reflects enterprise presence but not intensity of use within those organizations.
Claude Opus 4.6 — 12-hour time horizon, 1M context +

Anthropic releases Opus 4.6 with 1M token context window (beta), 128K output. METR measures 50% time horizon at ~719 minutes (~12 hours). 75.6% SWE-bench.

The ~12-hour time horizon meant agents could theoretically sustain productive work across an entire workday without human intervention. Combined with the 1M context window, this enabled agents to hold entire codebases in context while working on complex, multi-file tasks. The 99.9th percentile Claude Code turn duration nearly doubled to 45+ minutes.
GPT-5.3-Codex + Codex desktop app +

OpenAI releases GPT-5.3-Codex and desktop app for managing long-running agents. Demo: single session runs 25 hours, generates 30K lines of code.

The 25-hour demonstration — building a complete design tool with collaboration, layers, and export — was the most dramatic showcase of long-horizon capability. The agent performed validation (lint, type checking, tests, builds) after each milestone and auto-repaired failures. By March 2026, Codex had 2M+ weekly active users.
Cursor hits $2B ARR — doubles in 3 months +

Revenue doubles from $1B (Oct 2025) to $2B (Feb 2026). Enterprise customers now ~60% of revenue. Reportedly raising at $50B valuation.

Cursor shipped BugBot Autofix (cloud agents auto-fixing PRs, 35% merge rate), Cursor Automations (always-on agents triggered by code changes or Slack), and up to 8 parallel background agents on isolated VMs. Leadership explicitly called these "another step change in that progression."
Codex Security research preview +

OpenAI extends its Codex agent platform into security — automated vulnerability detection and remediation.

Security was a natural extension of the agentic paradigm: if agents can write code, they can also analyze it for vulnerabilities, propose fixes, and verify patches. This addressed one of the key concerns around AI coding: that increased code volume without proportional security review creates risk.
Claude Computer Use + Dispatch +

Anthropic ships phone-to-desktop dispatch: users message Claude to assign tasks that run autonomously on their computer.

This blurred the line between "coding agent" and "computer agent" — the same long-horizon execution architecture that powered Claude Code could now interact with any desktop application. Tasks could be assigned remotely and executed without the user being present at the machine.
Gartner: 75% of enterprise engineers to use AI by 2028 +

Gartner's forward-looking forecast — up from <10% in early 2023 and 63% of organizations piloting/deploying in Q3 2023.

This trajectory — <10% → 63% piloting → 75% forecast — illustrates the enterprise adoption curve lagging but following the individual developer curve. The constraint is procurement, policy, and compliance, not developer willingness.