Every AI Coding Agent Compared

Six months ago, Cursor was a code editor, Claude Code was a terminal assistant, and Codex was a cloud sandbox. Today, Cursor hit $2B in annualized revenue with over a million paying customers. Claude Code ships scheduled tasks, computer use, and remote control across your devices. Codex runs on GPT-5.4 with a desktop app and a model that generates code at 1,000+ tokens per second. Windsurf survived an acquisition, the departure of its founders, and a pricing overhaul. It still crossed a million active users.

Most comparison posts are written by the companies that make these tools, or sponsored by one of them. We built promptarc, a session viewer that parses every agent's local format. We've spent months inside their data. Here's what we actually learned.

A few numbers for context: 73% of developers now use AI coding tools daily. 41% of new code on GitHub is AI-assisted. The average developer uses 2 to 3 tools at the same time, not because they can't decide, but because different tools are better at different things.

The Taxonomy

Not all agents are the same shape. Three categories:

CLI-Native: Claude Code, Codex CLI, Aider, OpenCode. You control the environment. The agent operates in your terminal, your filesystem, your git repo. Best for complex multi-step tasks and CI/CD.

IDE-Integrated: Cursor, Windsurf, Kiro, Copilot, Cline, Continue. The agent lives inside your editor. Context is automatic. Best for daily coding with autocomplete + agent in one window.

Cloud/Async: Cursor Background Agents, Codex Cloud, Jules, Copilot Coding Agent. You assign a task, walk away, come back to a PR. Best for parallelizable and overnight batch work.

These categories are blurring fast. Claude Code now has cloud auto-fix. Cursor has a CLI. Codex has a CLI, a desktop app, and cloud tasks. Most serious developers use two or three tools across categories.

The Big 5

Claude Code

The autonomous terminal agent

In Q1 2026, Anthropic shipped 30+ releases in five weeks, transforming Claude Code into a full autonomous agent platform. The highlight reel: 1M token context window (Opus 4.6) with no surcharge on Max/Team/Enterprise. Scheduled tasks running on a cron-like schedule in Anthropic's cloud. Auto mode that reduces supervision by letting an AI classifier approve safe tool calls. Remote control lets you start on your laptop and continue on your phone. Computer use means the agent can open apps, click, navigate, and fill forms, not just run terminal commands. Cloud auto-fix follows your PRs, fixes CI failures, and addresses review comments without you lifting a finger.

The extensibility stack is the deepest of any CLI tool: ~101 official plugins, a 6-tier permission model (the most granular of any agent), CLAUDE.md project conventions, custom slash commands, and a skills system for packaging and sharing agent capabilities.

Developer sentiment tells the story: 46% "most loved" in the Pragmatic Engineer survey (15K devs), 91% CSAT and NPS of 54 in JetBrains' AI Pulse survey. No other tool comes close on satisfaction metrics.

The gap: Rate limits frustrate heavy users on the Pro plan. No built-in autocomplete (pair with Cursor or Windsurf). Session history is local JSONL with no built-in cross-session search.

Verdict: The most capable agent for complex, multi-step work. If you want an AI that stays active across devices, runs on a schedule, and operates autonomously, this is it. It expects you to be comfortable in a terminal.

Cursor

The agent orchestration platform

Cursor's numbers: $2B ARR, 2M+ users, over a million paying customers, trusted by over half of the Fortune 500. What started as a VS Code fork is now the most complete AI IDE.

The standout feature is background agents. You can run up to 8 in parallel, each in isolated Ubuntu cloud VMs with browser, desktop, and video recording. Supermaven autocomplete has an industry-leading acceptance rate with 100K token context for completions. Automations trigger agents from Slack, Linear, GitHub, PagerDuty, and schedules, making it the most comprehensive event-driven system among IDE tools. MCP Apps render interactive charts, diagrams, and whiteboards inside agent chats. Model flexibility across GPT-5.2, Claude Sonnet 4.5, Gemini 3 Pro, and more.

The gap: Credit pricing confuses some users because predicting monthly spend is hard. VS Code lock-in (no JetBrains). No persistent memory across sessions.

Verdict: The most complete AI IDE. If you want everything in one window (autocomplete, agent mode, background agents, plugin ecosystem), this is the default choice.

GitHub Copilot

The enterprise default

Copilot holds 42% market share, the largest install base. But only 9% call it "most loved". That gap between most used and most loved tells you everything about its position. It's the safe enterprise choice, not the exciting developer choice.

What it does well: assign a GitHub issue to Copilot and it autonomously creates a PR, self-reviews, and tags you when ready. Agentic code review gathers full project context and auto-generates fix PRs. Three-layer security scanning (CodeQL, secret scanning, Dependabot). Agent mode GA on VS Code and JetBrains. Org-wide custom instructions for enterprise governance.

The gap: Extensions only work in Ask mode, not Agent mode. No persistent memory. No automation triggers beyond GitHub events.

Verdict: If your company pays for GitHub Enterprise and compliance matters more than cutting-edge features, this is the safe bet. Not the most powerful, but the most integrated with where your code lives.

OpenAI Codex

The full-stack coding platform

Codex has evolved far beyond the original cloud sandbox. It's now a multi-surface platform: open-source CLI (Rust, Apache 2.0), macOS desktop app, cloud tasks, and IDE extension, all running on GPT-5.4 by default.

The speed story is real: GPT-5.3-Codex-Spark generates code at 1,000+ tokens/sec on Cerebras hardware. First-class plugins connect to Sentry, Datadog, Linear, Slack, Figma. GitHub Triggers automate responses to issues, PRs, and pushes. Kernel-level sandboxing uses Linux namespace isolation, the strongest security of any tool. Multi-agent v2 supports structured inter-agent messaging.

The gap: Smaller community than Claude Code or Cursor. macOS-only desktop. OpenAI models only. GitHub-only for background clone-and-push.

Verdict: The most complete offering from a single vendor: CLI, desktop, IDE extension, cloud tasks, fastest model. If you're in the OpenAI ecosystem with automation-first workflows, this is the play.

Windsurf

The enterprise wildcard

The most turbulent story in AI coding. Cognition acquired it for ~$250M after OpenAI's $3B deal fell through. Google paid $2.4B to license the technology and hire the founders. Three weeks later, Cognition laid off 30 employees and offered buyouts to ~200 remaining staff. The people who built Windsurf's core differentiation are no longer there.

Despite the turbulence: 1M+ active users, 4,000+ enterprise customers. The standout feature is Memories. Cascade autonomously creates persistent memories of your project context, coding patterns, and preferences across sessions. No other tool does this automatically. Broadest IDE coverage (VS Code, JetBrains, Vim, NeoVim, XCode, totaling 40+ plugins). Deepest compliance: SOC 2, HIPAA, FedRAMP, ITAR. Pricing overhauled March 2026: $20/mo (was $15), switched from credits to daily/weekly quotas.

The gap: Product direction uncertain under Cognition. No automation triggers. No plugin marketplace. Daily quotas frustrate heavy users. Memories are local-only with no cross-machine sync.

Verdict: The best compliance story and broadest IDE coverage. If you need HIPAA/FedRAMP, this is the only real option. But watch the changelog. The team turbulence is real.

The Challengers

Amazon Kiro

AWS's agentic IDE with a unique angle: spec-driven development. Natural language prompts become EARS requirements, then architecture, then task sequences. Uses Claude Sonnet 4.5 by default. Native MCP support. Free tier (50 credits), Pro at $20/mo, up to Power at $200/mo. Best for teams that want structured, requirements-first AI development.

OpenCode

Fully open-source (MIT), ~142K GitHub stars, 75+ LLM providers including local models via Ollama. Built by the SST team. Zero vendor lock-in. No memory, no cloud agents, no marketplace, but the community is building fast. Best for developers who want full transparency and model freedom.

Aider

100% open source, any LLM including local via Ollama, git-first philosophy where every AI change becomes a commit. Pioneered the architect/editor dual-model pattern. No MCP, no plugins, no memory, no cloud. Best for developers who want zero lock-in and think in git commits.

Cline

5M+ installs. Self-creating MCP tools: ask it to "add a tool" and it builds, installs, and configures an MCP server. Browser automation built-in. Plan/Act safety gate. Total cost transparency. Best for developers who want to see exactly what the AI spends.

Continue.dev

Open source with a unique angle: AI Checks as GitHub status checks (green/red on PRs). Headless CLI for CI/CD pipelines. Source-controlled quality standards. Best for teams that want AI in their merge gate.

Amazon Q Developer

Not a general-purpose coding agent. This is a migration and modernization tool. Java 8→17 upgrades, .NET porting, mainframe COBOL conversion, VMware workload migration. Deep AWS integration. Free tier with 50 agentic interactions/month. Best for enterprise cloud migrations on AWS.

Jules (Google)

Fire-and-forget: assign a task, come back to a PR. Gemini 2.5 Pro (free tier) or Gemini 3 Pro (Pro/Ultra). Runs in Google Cloud VMs. Jules Tools CLI and public API. Best for well-defined background tasks.

Comparison Tables

Capabilities

* Claude Code has no built-in autocomplete. Pair with Cursor or Windsurf for tab completion.

Pricing (April 2026)

What We Learned

We built a tool that reads the session formats of every agent on this list. Here's what surprised us:

Every tool is converging on the same architecture. Agent + sandbox + cloud + PR output. The differentiators are shifting from "what can it do" to "how does it feel to use."

No tool has cross-session search. You can search within a session, but "find the session where I fixed that OAuth bug last month" doesn't exist anywhere natively. (This is what we're building at promptarc.)

The bottleneck moved from generation to review. Agents produce code fast. Reviewing what they did is now the slow part.

Memory is the next battleground. Windsurf has Memories. Claude Code has CLAUDE.md. Cursor has Rules and Notepads. But none share knowledge across a team automatically.

The "2–3 tool" pattern is real. Most productive developers use one IDE agent for daily work + one CLI agent for hard problems + occasionally a cloud agent for background tasks. That's not indecision. It's the optimal setup in 2026.

How to Choose

Start with your #1 priority:

• Maximum autonomy → Claude Code
• Best IDE experience → Cursor
• Compliance / regulated industry → Windsurf
• Zero vendor lock-in → OpenCode or Aider
• GitHub-native → Copilot
• Speed → Codex (Spark at 1K tokens/sec)
• Spec-driven → Kiro
• Cheapest → Aider (free, BYOK)
• Legacy migrations → Amazon Q
• Fire-and-forget → Cursor BG or Codex Cloud

Most developers end up with 2–3 tools. One IDE agent for daily work + one CLI agent for hard problems + one cloud agent for background tasks. That's not indecision. It's the optimal setup in 2026.

Last updated: April 2026. If we got something wrong → promptarc.dev/feedback