Warp 2.0: The First Agentic Terminal (Open-Sourced April 2026)

Warp 2.0 is the terminal that decided to also be a coding agent.

Launched June 24, 2025 and open-sourced April 28, 2026, it's the project the Warp team has been telegraphing for two years — a rebuild of their terminal app around a coding agent that lives inside the prompt itself. The headline numbers are real: 71% on SWE-bench Verified, #1 on Terminal-Bench, and an AGPL-3.0 open-source release on April 28, 2026 (Warp's announcement). The GitHub repo currently shows 56,000+ stars (github.com/warpdotdev/warp). The launch post calls it the first Agentic Development Environment (warp.dev/blog/reimagining-coding-agentic-development-environment).

This page covers what Warp 2.0 actually is, how the SWE-bench number holds up to scrutiny, how it fits next to Claude Code and Cursor, and how the same RuleSell ruleset works across all three.

What Warp 2.0 actually is

The simplest framing: Warp 2.0 collapses the boundary between your shell and your coding agent. In a normal workflow, you have a terminal running shell commands and a separate agent (Claude Code in a different pane, Cursor in another window, Codex CLI in a third). Each context switch costs attention, and the agent doesn't see what you just typed in the shell.

Warp 2.0 puts both in the same window. The prompt accepts shell commands and agent instructions in the same buffer. When you type git status it runs as shell. When you type find the file that imports useAuth and refactor the import to use the barrel export it dispatches to the agent. The agent has read access to your scrollback, so what you ran a minute ago is part of its context.

Warp's launch post describes the surface as "Code, Agents, Terminal, and Drive in a single app" — Drive being a built-in file-sharing layer for code snippets and command history across a team. That's a bigger product surface than just a terminal, and it's worth being clear-eyed about: Warp 2.0 is a developer tool platform with a terminal as its primary UI, not just a terminal that grew agent features.

How the SWE-bench number holds up

71% on SWE-bench Verified is a strong claim. For context: most published Claude Code numbers on the same benchmark sit in the 60s, and Anthropic's flagship Claude 4.7 Opus runs around 70% in some configurations. So 71% from Warp puts it in the top tier of published systems. (Anthropic's own SWE-bench results and various community reproductions are the standard comparison points.)

A few caveats worth carrying:

• Vendor-published, not independently reproduced. The 71% is from Warp's launch post; we have not seen an independent reproduction yet. SWE-bench numbers can swing 5-10 points based on harness configuration (see /topic/agent-harness-engineering — the same model can swing 25.7 points on the same benchmark depending on harness).
• Verified vs. full. Warp claims SWE-bench Verified — the curated 500-issue subset. Full SWE-bench is harder and typically scores lower for the same system. The numbers are not interchangeable.
• Terminal-Bench is Warp's home turf. The #1 Terminal-Bench score is significant but the benchmark is newer and less standardized than SWE-bench. Treat it as a directional signal, not a settled ranking.

What AGPL-3.0 means for adoption

In May 2026, Warp open-sourced the project under AGPL-3.0. This is a deliberate license choice with real consequences:

• AGPL-3.0 is strong copyleft. Any modification that's offered as a network service requires source release. This blocks competitors from forking Warp into a closed-source SaaS.
• It limits enterprise adoption in AGPL-restricted shops. Many large companies prohibit AGPL software for production use because of the network-service clause. Those teams can run Warp locally but can't easily embed it in their own products.
• It protects Warp's hosted business. The hosted Warp service (paid tier) is where the company earns; the AGPL release keeps the project open without enabling a closed competitor.

Warp 2.0 vs Claude Code vs Cursor

The clean comparison, with no marketing:

• Claude Code: a CLI you run inside whatever terminal you already use. Best when your existing terminal setup matters (custom prompt, tmux, zsh customizations). Single-vendor model story (Claude).
• Cursor: a VS Code fork with deep agent integration. Best for IDE-centric workflows where the file tree and tabs are how you navigate. Multi-model via Cursor's model selector.
• Warp 2.0: the terminal itself, agent built in. Best when you live in the shell and want zero context switch. Multi-model via configurable backend.

What matters for RuleSell creators: a well-written ruleset is portable across all three. The format differs (CLAUDE.md vs .cursorrules vs Warp's agent config), but the substance — coding standards, project conventions, do/don't lists, examples — is reusable. RuleSell's multi-tool catalog is built around exactly this portability: write once, sell to all three audiences, earn payouts from each.

Where this fails / what we don't know

• Independent SWE-bench reproduction. As of this writing, the 71% number is vendor-published. We have no independent run to confirm.
• Production stability at scale. Warp 2.0 launched June 2025; the AGPL release was April 28, 2026. Six weeks of public production data is not enough to assess long-term reliability or churn.
• The Drive surface's real adoption. "Drive" — the team file-sharing layer — is part of the launch story but has no public usage data. It might be a meaningful product surface or it might be a feature on the marketing page.
• AGPL friction in practice. Whether AGPL-3.0 will actually limit Warp 2.0's penetration in enterprise versus just being noise in the procurement conversation is unknown. We'll know in 6-12 months.

What to read next

• /topic/cli-everything-wave — the broader meta-trend Warp 2.0 sits inside
• /topic/csi-2026-synchronized-output — the terminal-rendering spec Warp and its competitors all use
• /topic/agent-harness-engineering — why the same model scores differently in different harnesses
• /topic/dual-wielding-codex-claude-code — multi-CLI workflow patterns
• /for/warp — RuleSell rulesets targeted at Warp 2.0 users
• /for/claude-code — the comparable Claude Code catalog

Sources

• Warp blog, "Reimagining coding: agentic development environment" — warp.dev/blog. The launch post, SWE-bench claim, ADE framing.
• DEV Community, "Every AI Coding CLI in 2026: The Complete Map (30+ Tools Compared)" — dev.to. Warp's placement in the broader CLI landscape.
• Pinggy, "Top CLI Based AI Coding Agents" — pinggy.io/blog. Comparison context for Warp vs Claude Code vs Codex CLI vs Aider.
• SWE-bench project — swebench.com. Benchmark methodology and Verified subset definition.
• MindStudio, "Agent Harnesses Beat Model Upgrades" — mindstudio.ai/blog. 25.7pp harness-driven swing reference.
• AGPL-3.0 license text — gnu.org/licenses/agpl-3.0. Copyleft and network-service clause.