Warp 2.0

Warp 2.0 is the generation of Warp’s developer tooling launched on June 24, 2025 as an Agentic Development Environment. It combined coding, multiple-agent management, a terminal and shared team context in one application. Founder and CEO Zach Lloyd positioned the change around development by prompt: developers would assign tasks, review results and steer agents across coding, debugging and deployment.[2]

Warp 2.0’s universal input accepted both terminal commands and agent prompts. Agents located relevant files using grep, glob and codebase embeddings, with a dedicated planning mode using reasoning models such as o3. They could work across multiple repositories within one conversation, while developers could edit generated diffs directly in Warp. A management interface showed running agents’ status and notified users when an agent finished or needed help.[2]

Warp Drive supplied shared context for people and agents: rules, MCP configuration, commands, notebooks, environment variables and prompts. This made team conventions and saved operational knowledge available within agent tasks rather than requiring each developer to assemble the same context separately.[2]

Warp initially kept agent execution local. Lloyd explained that cloud agents removed humans from the loop too early for the technology’s capabilities at the time. Developers could configure approval requirements for file reads, commands, diffs and MCP access, establish command allowlists and denylists, and pause agents to redirect them. Warp also retained ordinary terminal use and provided a toggle to disable AI features.[2]

Warp used its own agents during development. At launch, Lloyd reported that Warp’s Rust codebase exceeded one million lines and that the new universal input had been built almost entirely by Warp’s agents. The claim concerned that interface, rather than the whole application.[2]

The 2025 agent environment built on an earlier terminal product and its collaboration tools. Warp 1.0 already had an input editor and block-based output; Warp Drive had introduced shared command workflows two years before the Warp 2.0 launch.[2][4]

Warp was founded in 2020 and partnered with Sequoia Capital in 2022. On June 21, 2023, Warp announced a $50 million Series B led by Sequoia, with Andrew Reed joining its board. The announcement named continued participation from Dylan Field, GV, Neo, BoxGroup, Jeff Weiner and Marc Benioff, alongside new angel investors Sam Altman and Tobi Lutke.[3][4]

Warp launched Drive alongside the financing announcement. Its initial release stored commands as named, described, parameterized workflows that developers could annotate, share and execute on demand. Updates to shared workflows synchronized to everyone with access. That release included workflows only; the broader store of notebooks, rules and agent context described in 2025 came later.[4][2]

On September 3, 2025, Warp launched Warp Code, a feature suite extending the coding capability introduced in June. Its emphasis was “agent steering”: reviewing changes as agents worked, requesting modifications against specific lines and editing diffs by hand. A lightweight native editor added tabbed file viewing, syntax highlighting, find and replace, and a file tree.[6]

Project initialization could create a version-controlled WARP.md file, with support for AGENTS.md, CLAUDE.md and Cursor rules. Slash commands managed rules, MCP servers and saved prompts. Agent Profiles paired a model with permissions so developers could assign different defaults to different tasks.[6]

Warp reported launch scores of 52% on Terminal-Bench and 71% on SWE-bench Verified in June 2025. By September 2, it reported 75.8% on SWE-bench Verified using GPT-5. Warp attributed the increase to GPT-5 high-reasoning and application changes including to-do lists, file editing and long-conversation management. These were dated company-reported benchmark results.[2][6]

At the June 2025 launch, Warp reported more than half a million active terminal users. In its July 29 update, it reported almost two million agent launches daily, nearly 100 million generated lines of code weekly, and more than two million codebases synced daily. It also reported revenue growth of 15 times since the start of 2025 and 100% over the preceding month, without disclosing an absolute revenue figure.[2][5]

The September Warp Code announcement reported more than 150 million generated lines each week and acceptance of over 97% of generated code diffs. Agent launches, codebases and accepted diffs measured different forms of activity: they did not establish unique customer counts or code correctness.[6][5]

On February 10, 2026, Warp launched Oz for cloud coding-agent orchestration. Lloyd said internal agent usage had begun exceeding laptop capacity and that the team wanted scheduled jobs and applications built around agents. Oz provided a CLI, API, SDK and scheduler, with environments combining Docker containers, Git repositories and startup commands. Session links, audit trails, reviewable artifacts and handoff to local work supported tracking and intervention. Warp offered hosted infrastructure and an enterprise self-hosted option.[7]

Lloyd reported that Oz was writing 60% of Warp’s internal pull requests at launch. He used it to recreate mermaid.js diagram functionality for Warp’s custom Rust UI framework, launching one agent per diagram type and using Computer Use to compare results with canonical Mermaid output. The feature was still under development.[7]

An unnamed Warp engineer also built an Oz fraud bot that ran every eight hours and proposed pull requests to block suspicious usage. Lloyd reported that one morning run identified and wrote PRs targeting nearly $60,000 of fraudulent Warp usage; this was a claim about usage targeted by countermeasures, rather than realized cash savings.[7]

On May 20, 2026, Warp extended bring-your-own-key access to its Free plan and added custom endpoints compatible with the OpenAI Chat Completions API. Developers could supply OpenAI, Anthropic or Google keys, or connect a router, internal gateway or hosted inference endpoint. Lloyd described this as part of making Warp more customizable following the recent open-sourcing of its client.[8]

Under the announced terms, Free-plan BYOK required login. Individuals and companies with ten people or fewer incurred no Warp credits for BYOK or compatible custom inference. Larger companies needed Business or Enterprise plans and paid platform credits based on agents’ active work on Warp’s managed systems, separately from the inference they supplied.[8]

Warp introduced Warp Factories in closed beta on August 18, 2026. It chose an infrastructure model: organizations would define and customize their own cloud workflows for triage, specification, implementation, review and verification, with humans involved at key decisions. Factory definitions lived in version-controlled code and specified repositories, agents, skills, MCPs and permissions. Factories supported different models and external harnesses, including Claude Code and Codex.[9]

Factories exposed throughput, cost and quality metrics, custom scorers and repeatable benchmarks comparing fixed tasks across configurations. With self-improvement enabled, observer agents scored runs and proposed PRs changing models, harnesses, context or skills. Factory MCP let local coding agents submit work, inspect status and pull work back for local iteration. Warp reported automating about 30% of its internal tasks through factories in August—a different measure from Oz’s earlier share of PR authorship.[9][7]