Industry report
Last updated: June 2026
State of Coding Agents
Mid-2026
A data-driven overview of the 26 AI coding agents shaping how software gets built: who makes them, where they run, how they interoperate, and where the ecosystem is heading.26Agents indexedAcross the landscape
25VendorsxAI owns two agents
13 / 13Open / ClosedAn even split
100%Support AGENTS.mdUniversal standard
96%Support MCPNear-universal
13New in 2025The breakout year
The Agents
26 agents, 26 vendors
The full field of AI coding agents tracked in this report, from frontier labs to independent startups. Hover to pause; tap any logo to visit its vendor.Timeline
The 2025 explosion
2025 was the breakout year. Half of all agents indexed (13 of 26) launched in a single year, driven by terminal agents and big-tech entrants.2025 alone produced 13 new agents, as many as 2021 through 2024 combined. The pace has not slowed: 2026 is already on the board with three more.
2021 (1)
GitHub Copilot launches as the first major AI coding assistant.Released this year- GitHub Copilot
2023 (4)
Cursor and Zed arrive as IDE-first approaches begin.Released this year- Factory
- Cursor
- Zed
- Aider
2024 (5)
Devin launches as the first "AI software engineer," an autonomous cloud agent.Released this year- Devin
- OpenHands
- Augment Code
- Warp
- Cline
2025 (13)
The breakout year. Claude Code accelerates the terminal-agent wave; OpenAI ships Codex; Google enters with Antigravity. Cognition AI acquires Windsurf.Released this year- Junie
- Trae
- Goose
- Claude Code
- VS Code (agent mode)
- Kilo Code
- Codex
- Amp
- opencode
- Kiro
- Qwen Code
- Antigravity
- Mistral Vibe
2026 (YTD) (3)
xAI and Xiaomi join as even more big-tech entrants arrive. xAI acquires Cursor.Released this year- Command Code
- Grok Build
- MiMo Code
Where They Live
The terminal is the universal surface
Agents can span multiple surfaces. 24 of 26 ship a terminal/CLI interface, but the trend is multi-surface: 18 of 26 (69%) run on two or more, and 14 span three or more.Terminal
24 · 92%
IDE Extension
11 · 42%
Web
11 · 42%
Desktop
9 · 35%
IDE (full)
7 · 27%
Multi-surface is the direction of travel. The terminal is the lowest common denominator, but the leading platforms increasingly span editor, web, and desktop too.
Terminal-only purists (6)CLI-first workflows, nothing else.
- Aider
- Amp
- Command Code
- Grok Build
- Mistral Vibe
- Qwen Code
Most surfaces, 4 categories (4)Meet developers wherever they are.
- Augment Code
- Claude Code
- Devin
- Factory
IDE-only (2)Editors that add agents natively.
- VS Code
- Zed
Standards Stack
A shared interoperability stack is emerging
Seven standards are converging across the ecosystem.AGENTS.md is universal, MCP and Skills are near-universal, and even newer standards like hooks are catching up fast.Rules (AGENTS.md)Repo-level instruction files agents read for project context and conventions
26/26 · 100%
SkillsReusable, packaged skill modules that extend agents with task-specific capabilities
25/26 · 96%
MCPModel Context Protocol, an open standard for connecting agents to external tools and data
25/26 · 96%
Custom Models / BYOKSupport for plugging in custom, self-hosted, or bring-your-own-key model endpoints
22/26 · 85%
SubagentsSupport for spawning or orchestrating sub-agents to delegate tasks within workflows
22/26 · 85%
PluginsPlugin/extension systems for third-party integrations and custom tools
20/26 · 77%
HooksLifecycle hooks that trigger custom behavior in response to agent events
19/26 · 73%
AGENTS.md is the de facto standard. 25 of 26 agents read
AGENTS.md for project context. 17 agents support all seven standards; the laggards are Aider (2/7), Warp (3/7), Zed (3/7), and VS Code (4/7).Interoperability
Standard convergence, implementation divergence
Agents agree onAGENTS.md, but beyond that each reads its own sprawl of proprietary rule files. .agents/skills/ is winning as the shared path for skills, yet 25 agents collectively maintain 30 distinct skills directories.Rules file
AGENTS.md25 of 26Proprietary rules
.cursor/rules, .windsurfrules, …26 of 26 (each has its own)Skills directory
.agents/skills/15 of 26 (30 distinct paths)Claude rules compat
CLAUDE.md10 of 26 (cross-compatible)MCP config
.mcp.json3 (emerging)Everyone reads AGENTS.md — then diverges. 25 of 26 agents read
AGENTS.md for project context, but each also reads its own proprietary instruction files. .agents/skills/ is the closest thing to a shared skills path, yet the 25 agents that support skills collectively use 30 distinct directory paths. The standard is settled; the implementation is not.Open vs Closed
An exact 50/50 split
Thirteen agents are open source and thirteen are closed, a remarkably even balance for a fast-moving market.Open vs closed source
26agents
Open source 13 (50%)
Closed source 13 (50%)
OSS license breakdown
13OSS
Apache-2.0 6 (46%)
MIT 5 (38%)
AGPL-3.0 1 (8%)
GPL-3.0 1 (8%)
GitHub stars: open-source repos (June 2026)
VS Code
187k
opencode
180k
Claude Code
135k
Codex
94k
Zed
86k
OpenHands
79k
Cline
64k
Warp
63k
Goose
50k
Aider
47k
Qwen Code
26k
Kilo Code
25k
Mistral Vibe
5k
Open at the edges. Several closed-source agents still publish their CLI as open source (for example, Claude Code's repo on GitHub) while the cloud platform stays proprietary. The boundary between open and closed is blurrier than the 50/50 split suggests.
Big Tech vs Startups
Big tech ties with startups, 13 to 13
Every major AI lab now ships a coding agent (OpenAI, Anthropic, Google, xAI, Mistral, Amazon) alongside Chinese giants Alibaba, Xiaomi, and ByteDance. xAI owns two: Cursor and Grok Build.Big tech & major AI labs (13)
AntigravityGoogle
Claude CodeAnthropic
CodexOpenAI
CursorxAI
GitHub CopilotGitHub
Grok BuildxAI
JunieJetBrains
KiroAmazon
MiMo CodeXiaomi
Mistral VibeMistral AI
Qwen CodeAlibaba
TraeByteDance
VS CodeMicrosoft
Independent & startup (13)
AiderAider-AI
AmpAmp
Augment CodeAugment Code
ClineCline
Command CodeCommandCodeAI
DevinCognition AI
FactoryFactory
GooseAgentic AI Foundation (Linux Foundation)
Kilo CodeKilo Code
opencodeAnomaly
OpenHandsAll Hands AI
WarpWarp
ZedZed Industries
Coding agents are now essential infrastructure. The fact that every frontier lab has shipped one, and that three Chinese tech giants are all in, signals the category has moved from experiment to strategic priority.
What's Next
Seven trends shaping the next phase
Where the coding-agent ecosystem is heading, based on the patterns visible across all 26 agents tracked here.1The terminal won, then expanded
Nearly every agent started terminal-first, then expanded to IDE, web, and desktop. The terminal remains the lowest-common-denominator surface, with 24 of 26 agents shipping a CLI.2Standards are converging fast
AGENTS.md went from 0 to universal in ~18 months, and MCP is following the same curve. The interoperability stack of rules, skills, hooks, MCP, plugins, subagents, and custom models is becoming table stakes.3Multi-agent orchestration is the new frontier
22 of 26 agents now support subagents, the ability to spawn child agents for parallel work. This points toward workflows where multiple agents collaborate on complex tasks.4BYOK is the norm
22 of 26 agents support custom or bring-your-own-key models. Vendor lock-in on the model layer is weakening as developers expect to choose their own backends.5Every major AI lab has an agent
OpenAI (Codex), Anthropic (Claude Code), Google (Antigravity), xAI (Grok Build), Mistral (Vibe), plus Chinese labs Alibaba, Xiaomi, and ByteDance. Coding agents are now essential infrastructure.6Open source holds at 50%
The ecosystem maintains a healthy balance between open and closed. Many closed-source agents still open-source their CLI components while keeping the cloud platform proprietary.7The IDE wars are back
Cursor, Trae, Kiro, and Zed are all AI-native IDEs competing for attention, while VS Code and JetBrains defend their territory with agent integrations like Copilot and Junie.