Factory AI
Autonomous coding agents (Droids) for agent-native development

An autonomous AI software engineer that turns tickets into tested PRs
Cosine's agent is a genuinely autonomous ticket-to-PR tool with native Jira, Linear, and GitHub integration, a generous free tier, and a rare air-gapped Sovereign option for regulated buyers. Buyers should note that its well-known SWE-bench claims are self-reported and never officially verified, that the Genie brand was retired for the Lumen family in 2026 (some naming churn), that disclosed funding is modest, and that autonomous PRs still require human review.
Cosine's agent, originally Genie and now part of the Lumen model family, is an autonomous AI software engineer that takes a ticket from GitHub, Jira, or Linear and delivers tested pull requests, trained on human-engineer reasoning. It offers a generous free tier and a rare air-gapped Sovereign option for regulated buyers. Its SWE-bench benchmark is self-reported and unverified, and it is agent-and-PR-centric rather than an in-editor autocomplete tool.
Cosine is a UK-based, Y Combinator-backed company whose autonomous coding agent first gained attention as Genie in 2024. The core idea is a fully autonomous software engineer: it takes a task from an issue tracker such as GitHub, Jira, or Linear, then writes and tests code and opens a pull request, trained by observing how human engineers reason rather than only on final outcomes. It is an agent-and-PR-centric product rather than an in-editor autocomplete tool. An important nuance for 2026: Cosine retired the Genie branding and now ships the Lumen model family, comprising Scout (lightweight/on-device), Outpost, and Sovereign (air-gapped, for regulated industries). Genie is therefore largely a legacy 2024-2025 name, though the product is current and evolving. The Lumen Sovereign push positions Cosine as a UK sovereign AI effort, reported to be backed by major British institutions, aimed at organizations that cannot send code to US clouds. At its 2024 launch, Genie claimed a then-strong SWE-bench score of about 30 percent, but this was self-reported and never submitted to the official leaderboard, so it is not independently verified. Cosine offers a generous free plan (up to 80 tasks, up to 100 projects, unlimited users, becoming read-only afterward), a $20 per seat Hobby plan, a $200 per seat Professional plan, and custom enterprise and self-managed options, on flat pay-per-task pricing. Disclosed funding is modest, with a $2.5 million seed in August 2024. It suits teams wanting to offload well-specified tickets into automated PRs and regulated buyers needing air-gapped agentic coding, but its benchmarks are unverified and its ecosystem is smaller than incumbents.
Cosine's autonomous coding agent, launched as Genie and now part of the Lumen family, takes tickets from GitHub, Jira, or Linear and delivers tested pull requests. It offers a generous free tier, flat pay-per-task pricing, and a rare air-gapped Sovereign tier for regulated buyers. Its well-known SWE-bench claim is self-reported and unverified, and the Genie brand was retired in 2026. It suits teams offloading well-specified tickets, not in-editor autocomplete.
Cosine is a UK-based (London), Y Combinator-backed company building an autonomous AI software engineer. Its agent launched as Genie in 2024 and, in 2026, was rebranded into the Lumen model family (Scout, Outpost, Sovereign).
The Lumen Sovereign effort positions Cosine as a UK sovereign AI offering, reported to be backed by major British institutions, targeting regulated and government buyers who cannot send code to US clouds.
The agent takes a task from GitHub, Jira, or Linear and autonomously writes and tests code, then opens a pull request, trained by observing human software-engineer reasoning. The Lumen family spans Scout (lightweight/on-device), Outpost, and Sovereign (air-gapped).
It integrates natively with issue trackers, supports many programming languages, and is oriented around agentic PR delivery rather than in-editor completion.
Engineering teams wanting to automate routine, well-specified ticket-to-PR work, plus regulated and government organizations needing air-gapped, sovereign agentic coding.
Software engineers and teams delegating well-specified tickets to an autonomous agent for tested PRs.
Engineering managers and platform leaders, plus procurement at regulated organizations evaluating sovereign options.
Senior engineers, security and compliance stakeholders, and DevOps teams.
A team or regulated organization wanting to automate routine, well-specified engineering tickets into tested pull requests, potentially in an air-gapped or sovereign environment.
Cosine raised a $2.5 million seed round in August 2024, led by SOMA and Uphonest Capital, with participation from Lakestar and Focal. Some aggregators cite a larger total around $8 million, but only the $2.5 million seed is clearly documented.
Genie was Cosine's autonomous coding agent, launched in 2024. In 2026 Cosine retired the Genie branding in favor of its Lumen model family (Scout, Outpost, Sovereign), so the product is current but rebranded.
It takes a task from GitHub, Jira, or Linear, then autonomously writes and tests code and opens a pull request, aiming to work like a human software engineer rather than just suggesting completions.
No. Genie's roughly 30 percent SWE-bench score from 2024 was self-reported and never submitted to the official leaderboard, so it is not independently verified.
Yes. The free plan allows up to 80 tasks with up to 100 projects and unlimited users, with no credit card required, after which access becomes read-only.
Yes. The Lumen Sovereign tier is designed for air-gapped, regulated deployments, positioned as a UK sovereign AI offering for organizations that cannot use US clouds.
Side-by-side pages for pricing, features, and best-fit use cases.
Autonomous coding agents (Droids) for agent-native development
Affordable, IDE-native agentic coding across the full dev lifecycle
AI coding agent built for JetBrains IDEs
Enterprise AI coding platform built for large, complex codebases