Mechanics that Travel
Vendors are not copying MCP or Plan Mode as brand names. They are copying machinery that already proved it can change a behavior that matters for agent products: trust enough to delegate, lower the cost of letting the agent cook, make work pick-up-able, and extend the product without rewriting the core.
Delegate with trust
People hand an agent more important work when they can inspect, constrain, and undo it.
Propose, then approve
The interesting part is not a Plan button. Someone proposes a piece of work, the work becomes visible before it becomes a change, feedback attaches to the proposal, and approval is what moves state. That is the pull-request mechanic wearing an agent costume. It lowers the cost of letting an agent cook because the human is reviewing a plan, not an already-mutated repo. Every tool that copied planning mode was copying that sequence, not a menu label.
- Planning Mode — Devin Desktop (v1.10.1, Jun 2025, first) → Cursor (v1.7, Sep 2025, +111 days) → Kiro CLI (v1.23.0, Dec 2025, +191 days) → Claude Code (v2.1.9, Jan 2026, +220 days) → GitHub Copilot CLI (v0.0.387, Jan 2026, +224 days) → OpenAI Codex CLI (rust-v0.90.0, Jan 2026, +230 days) → Antigravity CLI (v1.1.0, Jul 2026, +393 days)
- Todo / Task List — Claude Code (v0.2.93, Apr 2025, first) → GitHub Copilot CLI (v0.0.404, Feb 2026, +281 days) → OpenAI Codex CLI (rust-v0.133.0, May 2026, +386 days)
Undo as a first-class verb
Stories did not win because photos disappeared. They won because they lowered the pressure to create. Checkpoints do the same job for agents. If you can rewind the session, you do not have to be sure before you delegate. The mechanic is a cheap undo that makes ambitious runs feel reversible. Tools that shipped rewind were not adding a setting — they were buying the user permission to take a bigger risk.
- Checkpoints / Rewind — OpenAI Codex CLI (rust-v0.73.0, Dec 2025, first) → GitHub Copilot CLI (v0.0.393, Jan 2026, +39 days) → Kiro CLI (v2.7.0, Jun 2026, +179 days) → Claude Code (v2.1.191, Jun 2026, +191 days)
Permission envelopes
Delegation dies when the only options are 'trust everything' or 'approve every call'. An envelope — sandbox, allow-list, read-only, ask-once — lets people hand over more work without raising the blast radius. The mechanic is a fence you can see, not a terms-of-service paragraph. Once one vendor proved users would grant broader autonomy inside a fence, the rest of the category had to ship a fence too.
- Sandboxing / Permission Modes — Claude Code (v0.2.67, Apr 2025, first) → OpenAI Codex CLI (rust-v0.15.0, Aug 2025, +120 days) → GitHub Copilot CLI (v0.0.381, Jan 2026, +279 days) → Cursor (v2.5, Feb 2026, +313 days) → Antigravity CLI (v1.0.5, Jun 2026, +420 days) → Devin Desktop (v3.2.16, Jun 2026, +432 days) → Kiro CLI (v2.17.0, Aug 2026, +489 days)
Show the thinking
Verification gets more likely when the reasoning is inspectable. Visible thinking is not a party trick — it is a way to check the agent's work without standing up a second reviewer. People trust a long, consequential run more when they can watch the argument form. That is why extended-thinking toggles traveled: they turn a black box into something you can audit while it is still running.
- Extended Thinking — Devin Desktop (v1.3.9, Feb 2025, first) → Claude Code (v0.2.44, Mar 2025, +18 days) → OpenAI Codex CLI (rust-v0.4.0, Jul 2025, +134 days) → GitHub Copilot CLI (v0.0.375, Jan 2026, +316 days) → Kiro CLI (v2.6.0, Jun 2026, +465 days) → Antigravity CLI (v1.1.5, Jul 2026, +511 days)
Reduce commit anxiety
Unfinished or agent-proposed work is cheap to share and cheap to revert, so people actually let the agent run.
Propose, then approve
The interesting part is not a Plan button. Someone proposes a piece of work, the work becomes visible before it becomes a change, feedback attaches to the proposal, and approval is what moves state. That is the pull-request mechanic wearing an agent costume. It lowers the cost of letting an agent cook because the human is reviewing a plan, not an already-mutated repo. Every tool that copied planning mode was copying that sequence, not a menu label.
- Planning Mode — Devin Desktop (v1.10.1, Jun 2025, first) → Cursor (v1.7, Sep 2025, +111 days) → Kiro CLI (v1.23.0, Dec 2025, +191 days) → Claude Code (v2.1.9, Jan 2026, +220 days) → GitHub Copilot CLI (v0.0.387, Jan 2026, +224 days) → OpenAI Codex CLI (rust-v0.90.0, Jan 2026, +230 days) → Antigravity CLI (v1.1.0, Jul 2026, +393 days)
- Todo / Task List — Claude Code (v0.2.93, Apr 2025, first) → GitHub Copilot CLI (v0.0.404, Feb 2026, +281 days) → OpenAI Codex CLI (rust-v0.133.0, May 2026, +386 days)
Undo as a first-class verb
Stories did not win because photos disappeared. They won because they lowered the pressure to create. Checkpoints do the same job for agents. If you can rewind the session, you do not have to be sure before you delegate. The mechanic is a cheap undo that makes ambitious runs feel reversible. Tools that shipped rewind were not adding a setting — they were buying the user permission to take a bigger risk.
- Checkpoints / Rewind — OpenAI Codex CLI (rust-v0.73.0, Dec 2025, first) → GitHub Copilot CLI (v0.0.393, Jan 2026, +39 days) → Kiro CLI (v2.7.0, Jun 2026, +179 days) → Claude Code (v2.1.191, Jun 2026, +191 days)
Make work multiplayer
Work becomes visible, reviewable, and pick-up-able by a human or another agent without a meeting.
Spawn a worker
A pull request makes work visible to other people. Subagents and background tasks make work visible to other workers. The human stays on one thread; the extra labor happens off to the side, with its own state, and reports back. The mechanic is fan-out without forcing the user to become a process manager. Once Cursor and Claude Code showed that parallel agents felt like more throughput rather than more chaos, every rival had to answer it.
- Subagents / Parallel Agents — Claude Code (v0.2.74, Apr 2025, first) → Devin Desktop (v1.12.20, Oct 2025, +180 days) → Kiro CLI (v1.23.0, Dec 2025, +243 days) → GitHub Copilot CLI (v0.0.374, Jan 2026, +259 days) → Cursor (v2.4, Jan 2026, +278 days) → OpenAI Codex CLI (rust-v0.90.0, Jan 2026, +282 days) → Gemini CLI (v0.36.0, Apr 2026, +348 days) → Antigravity CLI (v1.1.6, Jul 2026, +461 days)
- Background Tasks — Claude Code (v1.0.71, Aug 2025, first) → GitHub Copilot CLI (v0.0.353, Oct 2025, +82 days) → Devin Desktop (v2.0.44, Apr 2026, +250 days) → Cursor (v3.2, Apr 2026, +259 days) → Antigravity CLI (v1.0.15, Jul 2026, +328 days) → Kiro CLI (v2.17.0, Aug 2026, +369 days)
Extend without rewriting
Third parties add capability so the product compounds instead of waiting on the vendor's roadmap.
Protocol, not plugin
A plugin marketplace is a vendor's garden. A protocol is a pipe anyone can attach to. MCP traveled because it let tools inherit the rest of the internet's connectors instead of rebuilding each one. The mechanic is interoperability as the product: once users could plug the same server into two agents, staying on a tool that could not speak the protocol started to feel like a tax. That is network-density machinery, which is why it copied faster than most features on this list.
- MCP Support — Devin Desktop (v1.3.3, Feb 2025, first) → OpenAI Codex CLI (rust-v0.9.0, Jul 2025, +160 days) → Kiro CLI (v1.23.0, Dec 2025, +308 days) → Gemini CLI (v0.23.0, Jan 2026, +328 days) → Cursor (v2.6, Mar 2026, +383 days) → GitHub Copilot CLI (v0.0.421, Mar 2026, +384 days)
Packaged recipes
Custom slash commands, skills, and lifecycle hooks are not power-user chrome. They let a behavior get written once and reused without opening the agent's internals. Contribution happens as a byproduct of solving your own problem, then sharing the recipe. The mechanic increases contribution without a giant reward — the same reason Stories and sticker packs spread. Tools that shipped a skill, command, or hook format were copying a way to let the crowd extend the product.
- Skills / Plugins — Claude Code (v2.0.12, Oct 2025, first) → OpenAI Codex CLI (rust-v0.69.0, Dec 2025, +62 days) → Devin Desktop (v1.13.6, Jan 2026, +95 days) → Gemini CLI (v0.25.0, Jan 2026, +104 days) → GitHub Copilot CLI (v0.0.389, Jan 2026, +104 days) → Cursor (v2.5, Feb 2026, +130 days) → Antigravity CLI (v1.0.9, Jun 2026, +250 days)
- Custom Slash Commands — Claude Code (v0.2.31, Mar 2025, first) → Devin Desktop (v1.8.2, May 2025, +61 days) → OpenAI Codex CLI (rust-v0.23.0, Aug 2025, +168 days) → Cursor (v1.6, Sep 2025, +190 days) → GitHub Copilot CLI (v0.0.329, Sep 2025, +208 days) → Gemini CLI (v0.28.0, Feb 2026, +342 days) → Kiro CLI (v2.7.0, Jun 2026, +464 days) → Antigravity CLI (v1.1.3, Jul 2026, +497 days)
- Hooks — Claude Code (v1.0.38, Jun 2025, first) → Cursor (v1.7, Sep 2025, +90 days) → Devin Desktop (v1.12.41, Dec 2025, +162 days) → Gemini CLI (v0.24.0, Jan 2026, +198 days) → GitHub Copilot CLI (v0.0.402, Feb 2026, +218 days) → OpenAI Codex CLI (rust-v0.130.0, May 2026, +312 days) → Kiro CLI (v2.13.0, Jul 2026, +381 days)
Keep context alive
The next session does not start from zero — the project, the web, and the user's intent survive.
Persistent instructions
AGENTS.md and CLAUDE.md are not documentation features. They are a way for work to leave enough context behind that the next session — or the next person — can pick it up without another meeting. The mechanic is memory that lives in the repo, not in a chat transcript that dies when you close the tab. Once one tool taught users to keep a file the agent always reads, every other tool had to honor the same file or lose the project's accumulated taste.
- Project Memory Files — Devin Desktop (v1.1.0, Dec 2024, first) → Claude Code (v0.2.54, Mar 2025, +105 days) → GitHub Copilot CLI (v0.0.385, Jan 2026, +405 days) → OpenAI Codex CLI (rust-v0.100.0, Feb 2026, +429 days) → Antigravity CLI (v1.1.6, Jul 2026, +590 days) → Kiro CLI (v2.18.0, Aug 2026, +610 days)
Live world input
An agent trapped in the repo is guessing. Web fetch, screenshots, and voice are the same mechanic in different costumes: let the model see the world the user is already looking at. That collapses the 'paste more context' ritual and makes the next turn cheaper. Vision and search copied quickly because users felt the absence immediately — a coding agent that cannot look at the error screenshot or the live docs is doing half a job.
- Web Search / Fetch — Devin Desktop (v1.2.1, Jan 2025, first) → Claude Code (v0.2.53, Mar 2025, +64 days) → GitHub Copilot CLI (v0.0.374, Jan 2026, +351 days) → OpenAI Codex CLI (rust-v0.139.0, Jun 2026, +509 days)
- Image / Vision Input — Devin Desktop (v1.0.5, Nov 2024, first) → Claude Code (v0.2.59, Apr 2025, +126 days) → GitHub Copilot CLI (v0.0.333, Oct 2025, +310 days) → OpenAI Codex CLI (rust-v0.64.0, Dec 2025, +371 days) → Antigravity CLI (v1.0.14, Jun 2026, +580 days)
- Voice Input — Devin Desktop (v1.11.0, Jul 2025, first) → OpenAI Codex CLI (rust-v0.105.0, Feb 2026, +224 days) → Claude Code (v2.1.71, Mar 2026, +233 days) → Cursor (v3.1, Apr 2026, +270 days) → Kiro CLI (v2.18.0, Aug 2026, +392 days)
Meet the user in habitat
The agent shows up where work already happens: the editor, git, CI, and the login people already pay for.
Native habitat
The mechanic is not 'we built an IDE plugin'. It is showing up in the habitat where the work already lives so using the agent does not feel like switching products. Editor extensions, git/PR actions, CI headless mode, and 'log in with the plan you already pay for' are all the same move: remove a context switch. Consumer apps learned this with sharing sheets and camera rolls. Agent tools learned it by embedding in VS Code, JetBrains, GitHub, and the subscription people already have.
- VS Code Extension — Claude Code (v2.0.0, Sep 2025, first) → GitHub Copilot CLI (v0.0.409, Feb 2026, +136 days)
- JetBrains Support — Devin Desktop (v1.0.2, Nov 2024, first)
- Git / PR Integration — Devin Desktop (v1.6.1, Apr 2025, first) → Claude Code (v2.1.20, Jan 2026, +300 days) → GitHub Copilot CLI (v0.0.420, Feb 2026, +332 days) → OpenAI Codex CLI (rust-v0.129.0, May 2026, +401 days)
- Headless / Non-interactive Mode — Claude Code (v0.2.66, Apr 2025, first) → OpenAI Codex CLI (rust-v0.8.0, Jul 2025, +101 days) → GitHub Copilot CLI (v1.0.33, Apr 2026, +377 days) → Devin Desktop (v2.1.29, Apr 2026, +384 days) → Antigravity CLI (v1.1.8, Jul 2026, +475 days)
- Subscription Login (Pro/Max) — Claude Code (v0.2.96, May 2025, first)
How this is built
The behaviors and mechanics are a human-curated taxonomy — judgment about why a capability copied, not an automated notice that it did. First-ship dates come from the same release-note pipeline as Who Copies Whom?. See the methodology page.
Download the taxonomy (JSON) · Query the API · Who copies whom · Compare tools