Warp

Warp is a developer-tools company founded by Zach Lloyd whose products span a terminal, coding agents and cloud orchestration. Its Warp Factories infrastructure, announced in closed beta on August 18, 2026, lets organizations build automated software-development loops with agents handling triage, specification, implementation and review, while humans intervene at chosen decision points. That work followed Warp 2.0’s integrated agent environment and Oz’s cloud-agent platform.[1][2][5][7]

The foundation for these products was Warp’s earlier terminal work: a custom Rust interface, command-and-output interactions, and shared team knowledge. Its history begins before the current generation of coding agents.[10][2]

Warp’s initial prototype used Electron. In Agata Cieplik’s September 2021 account, poor performance prompted the team to rewrite it in Rust, using GPU rendering and a custom UI framework developed with support from Nathan Sobo. Shikhiu, Michelle and Aloke had joined Lloyd in the early development effort. Cieplik reported seven engineers at that time; summer interns Saumya and Matt contributed notifications for long-running commands and custom theming.[10]

Warp launched its macOS public beta on April 5, 2022, with text-editor-like command editing, completion suggestions, searchable history, tabs and split panes. It also announced $23 million in financing: a previously raised $6 million seed round led by GV and a $17 million Series A led by Figma co-founder and CEO Dylan Field. Windows and Linux versions were still future plans at that launch.[11]

On June 21, 2023, Warp launched Warp Drive for saving, annotating, sharing and executing command workflows. These were parameterized commands with names and descriptions; updates synchronized to developers with access. The initial release contained workflows, with notebooks and other artifacts described as future work. Alongside it, Warp announced a $50 million Series B led by Sequoia Capital, with Andrew Reed joining its board.[3]

Warp released Agent Mode on June 17, 2024, replacing its first-generation Warp AI chat panel. It turned natural-language requests into multi-step terminal workflows: requesting permission to run commands, reading their output as context and adjusting subsequent commands after errors. It could learn an unfamiliar CLI from its help text. Natural-language detection used a local classifier; sending a request to Warp AI required an explicit user action.[4]

Michelle Lim’s launch demonstration made the interaction concrete. A developer could attach a “port 3000 already taken” error and type “fix it.” Agent Mode would ask to run kill $(lsof -t -i:3000), taking the port number from the attached output. In another example, it checked which Node version manager was installed before proposing an upgrade.[4]

Warp 2.0, launched June 24, 2025, combined Code, Agents, Terminal and Drive in an Agentic Development Environment. A universal input accepted both prompts and terminal commands. Agents found relevant files through tools such as grep and glob and through codebase embeddings; developers could use planning mode, edit generated diffs directly and monitor multiple agents with status displays and notifications. Drive supplied shared commands, notebooks, environment variables and prompts, alongside centralized MCP configuration and rules.[2]

The June 2025 design emphasized local agents and configurable autonomy. Developers could control file reads, command execution, diff acceptance and MCP access, maintain command allowlists and denylists, and pause agents. Lloyd said Warp deliberately began locally because cloud agents, at that stage of the technology, removed humans from the loop too early. Users could also continue using Warp as a terminal and disable its AI features.[2]

At the Warp 2.0 launch, the company reported more than half a million active terminal users. On July 29, 2025, it reported almost two million agent launches daily, nearly 100 million generated lines of code weekly and a 97%+ acceptance rate for agent-suggested diffs. It also reported revenue growth of 15 times since the start of 2025 and 100% over the preceding month. These were company-reported usage and relative-growth measures; diff acceptance measured acceptance rather than verified correctness.[2][9]

Warp launched Oz on February 10, 2026. Lloyd attributed its development to internal demand for more agents than local machines could handle, repetitive-job automation and agent-powered applications for managing user issues. Oz supplied a CLI, API, SDK and scheduler, with session-sharing links, reviewable artifacts and handoff to local work. Cloud environments combined Docker containers, git repositories and startup commands, supporting multi-repository tasks on Warp-hosted infrastructure or an enterprise self-hosted option.[5]

Lloyd described using Oz to recreate Mermaid diagram functionality in Rust for Warp’s custom UI framework. He launched one agent per diagram type, with agents using Computer Use to compare their output against canonical Mermaid results. The feature was still under development. At the Oz launch, he also reported that Oz was writing 60% of Warp’s pull requests—a measure of PR authorship, rather than the share of all engineering work automated.[5]

On April 28, 2026, Warp open-sourced its client under an AGPL license and introduced an agent-first contribution process managed by Oz. OpenAI was the repository’s founding sponsor, and GPT models powered the agentic management workflows; contributors could use other coding agents too. Agents were intended to handle coding, planning and testing, while community members supplied ideas, direction and verification. Warp retained a guiding role and said public GitHub issues would become the source of truth for feature tracking.[6]

Lloyd’s stated primary reason for opening the client was to improve Warp faster through community supervision of agents. He identified specification and behavioral verification as bottlenecks and also framed the decision as a way to compete with better-funded closed-source products. Warp reported nearly a million active developers in that announcement, without defining the activity window or identifying them as paying customers. In July, Lloyd described the Warp repository as a test bed for Oz’s software-factory approach, while internal engineers continued improving the terminal.[6][1]

Warp Factories entered closed beta on August 18, 2026. Warp chose to provide infrastructure on which organizations could build their own factories. Factory definitions lived in version-controlled code, including repositories, agent roles, skills, MCPs and permissions. A foreman agent took work from integrations such as Slack, Linear or GitHub and dispatched subagents to move it through the workflow. Humans could steer work or bring it into a local environment; teams could adopt individual stages before automating a larger loop.[7]

Factories included queryable throughput, cost and quality metrics, evaluation scorers and optional self-improvement loops. Observer agents scored a portion of runs and proposed pull requests changing factory definitions—for example, the model, context or skills used. Warp also offered bring-your-own inference, hosting and data-storage options. In the August announcement, Lloyd reported that factories automated about 30% of internal tasks, a different measure from Oz’s earlier PR-writing percentage.[7]

On September 3, 2026, Warp launched Factory Benchmarks in early access. Saved prompts, conversations, pre- and post-run git states and artifacts allowed tasks to be replayed from their initial state. Benchmarks executed a tasks × configurations matrix, with rubric-based LLM judges scoring dimensions such as correctness, efficiency and cost. Teams could use the results to define classifier-driven model-routing rules. Warp reported a 63% cost reduction on selected task types without a quality impact, later describing about 63% lower internal cost per PR over the preceding couple of weeks.[8]