Ellipsis (software)

Ellipsis is a cloud service for configuring and running large language model agents for software-development work. It launched in 2023 as a GitHub application that reviewed pull requests, posted summaries and review comments, and responded to a developer mention by answering questions or proposing code changes.^[1][3] By April 2025, Ellipsis described a broader "agents as code" system in which teams define an agent's instructions, trigger, repository access, permissions, and budget in a version-controlled YAML file.^[10]

Hunter Brooks and Nick Bradford founded Ellipsis in 2023, and the company joined Y Combinator's Winter 2024 batch.^[1][2] The company announced a $2 million seed round in June 2024.^[5] Ellipsis should not be confused with similarly named software packages or with the three-dot punctuation mark.

History

Ellipsis announced its Y Combinator participation on December 28, 2023. The announcement identified Brooks and Bradford as the company's two engineers and described the initial product as a GitHub app that reviewed every pull request, flagged logical errors, security issues, and style-guide violations, and could run a requested change after a user tagged @ellipsis-dev.^[1] Y Combinator's company profile lists Ellipsis as founded in 2023 and as a member of its Winter 2024 batch.^[2]

The founders' Y Combinator launch biographies described Brooks as a former machine learning engineer at Amazon Web Services and Amazon Ads. Bradford was described as a former founding engineer at the code-migration company Grit, an interpretability researcher at Redwood Research, and an engineering manager at Hyperscience.^[2] These biographies explain the technical backgrounds presented at launch; they are not used here to infer either founder's later employment.

Brooks and Bradford introduced the review product on Hacker News on February 27, 2024. Their launch post said that Ellipsis appended a change summary to a pull-request description, checked the change against general and user-defined rules, produced another review after each new commit, and answered questions in a pull-request thread. The post described code generation as a public beta at that time.^[3] A second launch post on May 9 presented code changes through direct commits or separate pull requests as part of the product.^[4]

On June 19, 2024, Ellipsis announced a $2 million seed round. The company named Twenty Two Ventures, Four Cities Capital, Garage Capital, Amino Capital, Transpose Platform Management, Pioneer Fund, several Y Combinator leaders, and other investors as participants.^[5] The announcement date is the date of the public funding disclosure; it does not establish the legal closing date of the financing.

The product expanded beyond its original GitHub-only implementation. In a September 1, 2024 engineering post, Ellipsis said it had added GitLab support and described the different access-token, bot-account, group, and webhook setup required for GitLab installations.^[6] This makes a blanket description of Ellipsis as GitHub-only inaccurate for the period after that announcement.

Product and workflow

The original Ellipsis workflow centered on pull requests. A repository event started a review, and the service examined the change, posted a summary and selected findings, and repeated the review when the pull-request branch changed. Teams could express project-specific review rules in natural language. A developer could mention @ellipsis-dev to ask a question, request another review, or ask the service to make a change.^[1][3][4]

Ellipsis later announced workflows outside the pull-request review itself. In October 2024 it introduced implementation plans for GitHub and Linear issues. A user could request a plan with a mention, or configure automatic planning for newly opened issues. The company said the plan used repository context and prior pull requests and could ask clarifying questions when the task was ambiguous.^[7] A separate October 2024 release added daily, weekly, or monthly codebase-change reports delivered through Slack or email.^[8]

In an April 2025 product essay, Ellipsis described agents as YAML definitions stored in a repository. Under that model, an agent definition specifies its instructions, trigger, repositories, permissions, and spending limit. The company described scheduled triggers, event triggers from GitHub, Linear, Slack, or Sentry, mention triggers, and on-demand sessions started through a dashboard, API, or command-line interface. Each execution was described as a separate cloud session in an isolated sandbox, while the version-controlled definition remained the durable configuration.^[10] These are descriptions published by Ellipsis, not the results of an independent feature or security audit.

System design

Nick Bradford published a technical account of the service's architecture on January 30, 2025. In that account, events from the GitHub App passed through Hookdeck to a FastAPI application and then to a Hatchet workflow queue. A pull-request event caused a worker to clone the repository and run a review workflow. Bradford said the asynchronous design allowed the system to favor review accuracy over interactive response time.^[9]

The review system was decomposed into multiple smaller AI agents, rather than one prompt handling every task. Several comment generators ran in parallel and targeted different issue classes, such as custom-rule violations or duplicated code. Their draft comments then passed through filters for confidence, duplication, and logical correctness. The system attached code evidence to draft findings and used that evidence during filtering. User ratings and written replies were also retrieved to help suppress kinds of comments that had previously been judged unhelpful.^[9]

The same account described a code-search subagent shared by review, code generation, and codebase chat. It combined keyword and vector retrieval. One index split source code into functions and classes with tree-sitter, while another embedded model-generated file summaries for broader questions. An LLM-based classifier selected retrieved material using the current search trajectory instead of relying only on a cosine-similarity threshold. Pull requests were indexed separately so that agents could find earlier examples of related changes.^[9] This is a form of retrieval-augmented generation, but the post did not provide an independent comparison with other retrieval designs.

Bradford also described placing a language-server proxy in a sidecar container so agents could use operations such as go-to-definition and find-references. At the time of the post, Ellipsis used a mixture of GPT-4o and Claude models and was experimenting with OpenAI's o1 for some tasks. Those model names document the January 2025 implementation and should not be read as a permanent list of providers or versions.^[9]

For evaluation, the team described small task-specific benchmark sets, deterministic checks where outputs had an exact answer, LLM-based judges for less exact outputs, and trajectory auditing to diagnose failed agent steps. Bradford wrote that the team generally preferred prompt changes, examples, and smaller composable agents before fine-tuning.^[9] The publication describes an internal development process; it does not report a blinded or independently replicated product evaluation.

Security and data handling

Ellipsis announced on August 1, 2024 that it had completed a SOC 2 Type I examination. Its announcement said repositories were cloned into ephemeral sandboxes and discarded when a job completed, and directed readers to contact the company for the report.^[11] Because the public announcement does not include the auditor's report, it supports the fact that the company made the announcement but does not allow a reader to inspect the examination's scope, controls, exceptions, or report date.

The January 2025 architecture post said that the vector-search system stored chunk hashes, obfuscated location metadata, and embeddings rather than customer source-code text. It also described cloning a repository for each review workflow.^[9] The April 2025 platform description added isolated cloud sandboxes, repository-scoped credentials, per-session and trailing spending limits, and session records containing triggers, transcripts, tool calls, output, and cost.^[10] All of these claims come from the vendor. Organizations evaluating the service would need the applicable contract, security documentation, subprocessors, retention terms, and audit report to determine how the controls apply to their own repositories and data.

Adoption and evidence

The June 2024 funding announcement said more than 90 companies were using Ellipsis. It also said that, according to the company's data, teams using the service merged pull requests 13 percent faster than teams without AI code review.^[5] Ellipsis did not publish the sample, assignment method, uncertainty, analysis code, or a peer-reviewed study with that announcement. The 13 percent figure is therefore a company-reported observational claim, not evidence that installing Ellipsis causes the same improvement for another team.

The May 2024 Hacker News discussion provides public examples of early reactions but not a controlled evaluation. Some commenters reported useful findings, while others pointed to verbose summaries, incorrect suggestions, and a risk that an automated first pass could create false confidence. The founders themselves described the tool as supplementing rather than replacing human review.^[4] These comments are anecdotal and cannot establish an error rate or a general productivity effect.

Academic work available by July 2026 studied AI-assisted code review as a broader practice rather than independently testing Ellipsis. One mixed-methods study of more than 100 students examined an LLM reviewer integrated into GitHub pull requests and reported both engagement measures and the need for guidance against over-reliance.^[12] A separate research vision for agentic code review kept humans at decision gates across a proposed five-stage review workflow.^[13] These studies provide context for evaluating human oversight, but their findings do not validate Ellipsis's feature, accuracy, security, or productivity claims.

Related

CodeRabbit - Greptile - GitHub Copilot - AI code generation

References

1. ^Ellipsis Joins Y Combinator, Ellipsis, December 28, 2023
2. ^Ellipsis: AI code reviews and bug fixes, Y Combinator
3. ^Show HN: Ellipsis - Automatic pull request reviews, Hacker News, February 27, 2024
4. ^Show HN: Ellipsis - Automated PR reviews and bug fixes, Hacker News, May 9, 2024
5. ^Ellipsis Raises a $2M Seed Round, Ellipsis, June 19, 2024
6. ^How to Move Your GitHub App to GitLab, Ellipsis, September 1, 2024
7. ^Launch Alert: Implementation Plans for GitHub/Linear Issues, Ellipsis, October 10, 2024
8. ^Launch Alert: Codebase Reports, Ellipsis, October 17, 2024
9. ^Nick Bradford, How we built Ellipsis, January 30, 2025
10. ^The Future of Software Is the Agent, Ellipsis, April 25, 2025
11. ^Ellipsis Is Now SOC 2 Type I Certified, Ellipsis, August 1, 2024
12. ^Eduardo Oliveira et al., AI-Assisted Code Review as a Scaffold for Code Quality and Self-Regulated Learning: An Experience Report, 2026
13. ^Huseyin Ozgur Kamali et al., Rethinking Code Review in the Age of AI: A Vision for Agentic Code Review, 2026