Augment Code
Last reviewed
May 6, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 ยท 4,324 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 6, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 ยท 4,324 words
Add missing citations, update stale details, or suggest a clearer explanation.
Augment Code is an enterprise AI coding platform founded in 2022 and headquartered in the San Francisco Bay Area. The company positions itself as a provider of AI agents purpose-built for large, complex codebases. Its core product is a suite of coding agents and developer tools built around a proprietary Context Engine that indexes entire repositories to give AI models accurate, relevant information before they write a single line of code. In April 2024 the company emerged from stealth with $252 million in total funding. By early 2026, Augment Code's flagship agent, Auggie, held the top spot on Scale AI's SWE-bench Pro leaderboard with a 51.80% resolve rate, demonstrating that agent architecture matters as much as the underlying model.
Igor Ostrovsky and Guy Gur-Ari started working on the company in 2022. Ostrovsky had spent nearly seven years at Microsoft working on the research operating system Midori before joining Pure Storage as a founding engineer and eventually becoming chief architect. Gur-Ari came from Google, where he worked as an AI research scientist. The two built the initial prototype around the idea that developers working on large, real-world codebases needed something fundamentally different from generic autocomplete: they needed AI that understood how a codebase actually fit together.
To build out the executive team, they brought in Scott Dietzen as CEO. Dietzen had previously led Pure Storage through its IPO and spent time at Yahoo! and BEA Systems. Dion Almaer joined as VP of Product, drawing on experience at Google, Shopify, Mozilla, and Palm. The broader engineering team drew talent from Google, Meta, NVIDIA, Microsoft, Databricks, Snowflake, VMware, and Pure Storage.
The company operated in stealth through 2022 and 2023, publicly emerging on April 24, 2024, when it announced its $227 million Series B. At that point the company had roughly 50 employees and was targeting enterprise software teams working on codebases too large for generic AI tools to handle well.
Augment operated initially under the name Augment Inc. The product branding shifted to Augment Code as the company sharpened its focus on software engineering specifically, distinguishing it from broader "augmentation" framing.
Augment Code raised its initial $25 million Series A led by Sutter Hill Ventures before emerging from stealth. On April 24, 2024, the company announced a $227 million Series B at a $977 million post-money valuation, bringing total funding to $252 million.
The Series B attracted a broad investor group including Sutter Hill Ventures, Index Ventures, Innovation Endeavors, Lightspeed Venture Partners, and Meritech Capital. Eric Schmidt, the former Google CEO, also participated as an individual investor. Innovation Endeavors is Schmidt's venture fund, which explains the personal involvement alongside the fund.
The capital was designated for accelerating product development and expanding engineering, product, and go-to-market functions. At the time of the announcement, the company said it planned to roughly double its headcount by the end of 2024. By 2025 the company had grown to approximately 156 people and was reporting roughly $20 million in annual revenue.
| Round | Date | Amount | Valuation | Lead Investors |
|---|---|---|---|---|
| Series A | 2022-2023 | $25M | Undisclosed | Sutter Hill Ventures |
| Series B | April 2024 | $227M | $977M | Sutter Hill Ventures, Index Ventures, Lightspeed |
Auggie is Augment Code's primary AI coding agent. It is available as an IDE extension for Visual Studio Code and JetBrains products, as a command-line tool (Auggie CLI), and through the web-based Intent workspace. The agent uses Augment's proprietary Context Engine to understand a codebase before acting on instructions.
Where many coding agents treat the codebase as a text corpus and rely on keyword or embedding search to locate relevant files, Auggie's Context Engine builds a semantic index that maps architectural relationships: which services call which interfaces, which modules depend on shared utilities, how patterns repeat across a repository. When Auggie receives a task such as adding logging to payment requests, it traces execution paths from frontend through API layer through backend services and database handlers, then applies changes consistently using the team's existing code patterns rather than generic idioms.
Auggie supports a range of task types. In interactive sessions within an IDE, it maintains a task list for complex work and automatically writes session memory so context persists between sessions. For long-running or CI-driven workflows, it operates in automation mode through the CLI. The agent can spin up sub-agents for specialized work such as security audits or test writing, run them in parallel, and then verify results before returning control to the developer.
The agent integrates with MCP (Model Context Protocol) to connect with external tools including GitHub, Linear, and Jira, pulling in ticket context and issue history as part of task planning. It also supports tool permissions management, allowing teams to restrict what operations the agent can execute in production environments.
In April 2025 Augment Code released an open-source version of its SWE-bench agent, achieving 65.4% on SWE-bench Verified as the top published open-source score at the time. That agent combined Claude Sonnet 3.7 as the primary driver with OpenAI's o1 model as an ensemble component for candidate solution selection. The company open-sourced the implementation at augmentcode/augment-swebench-agent on GitHub to establish a public baseline and share methodology.
Augment Code distributes its primary developer-facing interface through IDE extensions rather than a standalone editor. This positions it differently from Cursor (code editor), which ships its own fork of VS Code as a full replacement.
The VS Code extension provides the full feature set: Auggie agent for multi-file autonomous tasks, chat for interactive Q&A and code review, code completions (described as "Next Edit" suggestions that guide developers through complex changes), and remote agent capabilities for cloud-based task execution. The extension indexes the local repository on first install and maintains a live semantic index as files change.
The JetBrains plugin covers the full JetBrains IDE family: IntelliJ IDEA, PyCharm, WebStorm, GoLand, Rider, DataGrip, and CLion. The JetBrains plugin initially shipped with chat and completions; full agent capabilities were added as the product matured. The plugin is listed on the JetBrains Marketplace.
Augment Code also supports Vim/Neovim and Emacs through its ACP (Agent Client Protocol) interface, which exposes the same agent and context capabilities to editors that support the protocol. Zed, a newer Rust-based editor, is also listed as a supported ACP client.
Code completions are available across all supported IDEs. In late 2025, Augment Code announced that code completions would be deprecated for the Indie, Standard Max, and Legacy plans as of March 31, 2026, shifting focus toward agent-driven workflows over single-line or block completions.
Auggie CLI brings the same agent and Context Engine capabilities to the terminal. It is installed via npm:
npm install -g @augmentcode/auggie
The CLI operates in two modes. Interactive mode streams output in real time and shows visual progress, suited for exploratory tasks where a developer watches and steers. Automation mode uses --print and --quiet flags to suppress interactive output, making it suitable for use in CI/CD pipelines.
Key CLI capabilities include parallel agent execution for running multiple sub-agents simultaneously, resumable sessions that preserve conversation history across terminal sessions, prompt enhancement that refines rough natural-language instructions before execution, and a /github-workflow command that generates GitHub Actions YAML for custom automation.
For CI/CD integration, Augment Code ships official GitHub Actions: augmentcode/describe-pr and augmentcode/review-pr. These actions attach context-aware comments directly to pull requests with full codebase understanding, not just analysis of the diff in isolation. The CLI also works with Jenkins and any other CI system that can execute shell commands.
Service accounts with dedicated API tokens allow production workloads to authenticate without personal developer credentials. Sessions can be shared and exported as markdown for documentation or handoffs.
The CLI supports the same MCP integrations as the IDE extension, connecting to GitHub, Linear, Jira, and custom MCP servers. This means an agent running in a CI pipeline has access to the same ticket context and repository history as a developer working interactively in VS Code.
The Context Engine is Augment Code's most technically differentiated component and the one the company credits for its benchmark results. It is a full search and indexing system for code, designed to provide AI models with the minimal set of relevant context needed to complete a task accurately, without flooding the model with irrelevant tokens.
The engine works by building a semantic index of the entire repository: files, functions, classes, interfaces, call graphs, import graphs, and dependency relationships. Rather than treating code as unstructured text, it understands that a TypeScript interface in one file has concrete implementations spread across a dozen service modules, and that changing the interface requires touching all of them.
When an agent receives a task, the Context Engine runs a retrieval step that identifies relevant code through multiple strategies: semantic similarity (embedding-based search across the codebase), structural relationship traversal (following call graphs and import chains), pattern matching against team conventions, and historical information from commit history. The engine ranks and deduplicates results, then compresses the final context window to remove redundancy without losing critical detail.
For very large codebases, the engine uses quantized vector search to reduce memory overhead. As described in Augment Code's engineering blog, naive vector indexing of a 100 million line codebase requires roughly 2 GB of memory and can take over 2 seconds per search query. By applying quantization techniques, the company reduced memory to approximately 250 MB and cut search latency to under 200 milliseconds, maintaining more than 99.9% fidelity to exact search results.
The Context Engine also incorporates commit history. A feature called Context Lineage indexes the full git history of a repository, allowing the agent to understand not just the current state of the code but why specific decisions were made. When a file contains a comment referencing a past architectural discussion or a function was rewritten to fix a subtle race condition, the Context Engine can surface that history as part of a task's context.
Augment Code describes the engine as a proprietary layer that differentiates the product even when the underlying language model is the same as a competitor's. The SWE-bench Pro result in February 2026 provided a controlled demonstration: Auggie, Cursor, and Claude Code all ran Claude Opus 4.5 on the same benchmark, yet Auggie solved 15 to 17 more problems than the other two. The company attributed the gap to the quality of context retrieval, which determined whether the agent located low-level utility functions critical for multi-file fixes.
Augment Code targets engineering organizations working on codebases with millions to hundreds of millions of lines of code. The company's engineering blog has documented scaling work for repositories at the 100 million line scale, and the product marketing specifies support for codebases up to 100 million lines and beyond.
For enterprise deployments, the platform provides:
Security and compliance: Augment Code achieved SOC 2 Type II attestation in July 2024 after a three-month audit covering April through July 2024. In August 2025 the company became the first AI coding assistant to achieve ISO/IEC 42001 certification, which covers AI management systems. The Standard and higher tiers include Customer-Managed Encryption Keys (CMEK), SIEM integration, data residency options, and ISO 42001 compliance. Enterprise plans add SSO/OIDC/SCIM support for identity federation.
Data privacy: Augment Code's paid plans do not use customer code to train AI models, a commitment that extends to third-party model providers. The platform is architected so that AI suggestions are informed only by code accessible to the specific engineer making the request, respecting existing access controls.
Multi-repository context: Enterprise codebases frequently span dozens of repositories across services, libraries, and internal packages. The Context Engine is designed to index and search across multiple repositories simultaneously, surfacing the correct internal library function or shared utility even when it lives in a different repo from the file being edited.
Organizational memory: The Cosmos platform, which entered public preview in May 2026 for MAX plan users, introduces a shared knowledge layer across an entire engineering organization. Rather than each developer's agent session starting from scratch, Cosmos maintains organizational context, team conventions, and accumulated understanding across all agents and all developers.
Code review automation: The platform includes a code review agent that posts inline comments on pull requests backed by full codebase context. This distinguishes it from tools that analyze only the diff: the Augment Code review agent understands how the changed code relates to the broader system, surfacing issues that only become visible with codebase-level understanding.
Customers documented in Augment Code's case studies include Pure Storage (2,000 engineers, targeting 3x velocity), Tekion (1,300+ engineers, reporting 50-100% productivity gains and over 90% automated test coverage), GoFundMe (400-person engineering team), Intercom (multi-agent workflows on large Ruby and JavaScript codebases), DXC Technology (timeline reduction from months to days for certain software delivery tasks), Quorum (feature timelines cut from three weeks to three days), and WEX (automating full software delivery lifecycle with connected AI agents). MongoDB, Spotify, Snyk, Webflow, Crypto.com, MoneyGram, Rubrik, and Pigment are among the companies shown as customers.
SWE-bench is a benchmark that tests AI systems on real GitHub issues from popular open-source projects. The system must navigate the codebase, understand the reported problem, and produce a patch that passes the project's regression tests without the agent having been shown the tests themselves. SWE-bench Verified is a curated subset of 500 problems verified by human annotators as valid and unambiguous. SWE-bench Pro, released in late 2025, is a harder variant that includes multi-language problems, multi-file changes (averaging 4.1 files and 107 lines changed per solution), and more diverse task types beyond simple test-failure reproduction.
Augment Code has maintained a visible presence on both benchmarks:
SWE-bench Verified (open-source agent, March 2025): Augment Code released an open-source agent that achieved 65.4% on SWE-bench Verified, the highest published open-source score at the time. The agent combined Claude Sonnet 3.7 as primary driver and OpenAI o1 as an ensemble component for majority-vote solution selection.
SWE-bench Verified (Auggie, mid-2025): Augment Code subsequently reported a 72.0% result on SWE-bench Verified using Claude Opus 4.6 and Auggie's production context engine. This score reflects single-attempt (pass@1) evaluation.
SWE-bench Pro (Auggie, February 2026): Auggie took the top spot on the SWE-bench Pro leaderboard with 51.80%, evaluated on 731 problems. The same benchmark had Cursor at 50.21%, Claude Code at 49.75%, OpenAI Codex at 46.47%, and the SWE-Agent baseline at 45.89%. All three top finishers (Auggie, Cursor, Claude Code) ran Claude Opus 4.5.
| Benchmark | Agent | Model | Score | Date |
|---|---|---|---|---|
| SWE-bench Verified | Augment OS agent | Claude Sonnet 3.7 + o1 | 65.4% | March 2025 |
| SWE-bench Verified | Auggie | Claude Opus 4.6 | 72.0% | Mid-2025 |
| SWE-bench Pro | Auggie | Claude Opus 4.5 | 51.80% | February 2026 |
| SWE-bench Pro | Cursor | Claude Opus 4.5 | 50.21% | February 2026 |
| SWE-bench Pro | Claude Code | Claude Opus 4.5 | 49.75% | February 2026 |
| SWE-bench Pro | OpenAI Codex | GPT-5.2-codex | 46.47% | February 2026 |
| SWE-bench Pro | SWE-Agent baseline | Claude Opus 4.5 | 45.89% | February 2026 |
The controlled nature of the February 2026 evaluation, where three agents ran the same model, is the most direct evidence for Augment Code's claim that agent architecture and context retrieval quality produce meaningful performance differences independent of model choice. The 15-problem gap between Auggie and Claude Code on 731 problems is not explained by model selection, since both ran Opus 4.5.
Augment Code shifted from a per-message pricing model to a credit-based system effective October 20, 2025. Credits reflect the actual computational cost of each request, scaling with prompt length, token usage, and model selection. The change attracted criticism from some users who reported cost increases of up to ten times previous rates, particularly for heavy agentic workflows.
| Plan | Price | Monthly Credits | Users | Notable inclusions |
|---|---|---|---|---|
| Indie | $20/month | 40,000 | 1 | Context Engine, Agent, Chat, MCP tools, SOC 2 Type II |
| Standard | $60/month per developer | 130,000 (pooled) | Up to 20 | All Indie features, advanced analytics, user allowlist, CMEK, ISO 42001, SIEM integration, data residency |
| Max | $200/month per developer | 450,000 (pooled) | Up to 20 | All Standard features plus Cosmos access |
| Enterprise | Custom | Custom | Unlimited | All features, SSO/OIDC/SCIM, dedicated support, annual volume discounts, custom top-up rates |
Credits are pooled at the team level on Standard, Max, and Enterprise plans, so heavy and light users within the same team share a common bucket. Auto top-up is available at $15 per 24,000 additional credits. As a reference point, the company has published that a small task on Claude Sonnet 4.5 costs approximately 293 credits, a medium task approximately 860 credits, and a complex agentic task approximately 4,261 credits.
A free trial is available for new users who add a valid payment method. Code completions are being deprecated for Indie, Standard Max, and Legacy plans as of March 31, 2026, reflecting the company's shift toward agent-based workflows.
Prism, a model-routing option added in May 2026, reduces credit consumption for routine tasks by automatically selecting cheaper models when the quality difference is negligible. The company estimates Prism delivers 20-30% cost savings versus always using frontier models, at comparable quality on its internal benchmarks.
The AI coding assistant market in 2025 and 2026 includes tools that differ significantly in their architecture: standalone IDEs (Cursor (code editor), Windsurf), IDE extensions (Cline, GitHub Copilot, Augment Code), and CLI-first agents (Claude Code). Augment Code's distinguishing claims center on codebase scale, enterprise compliance, and the Context Engine's retrieval quality.
| Tool | Type | Context approach | Max codebase scale | Pricing | SOC 2 Type II | SWE-bench Pro |
|---|---|---|---|---|---|---|
| Augment Code | IDE extension + CLI | Semantic index, up to 400K+ files | 100M+ lines documented | $20-200/mo per developer | Yes (July 2024) | 51.80% (Feb 2026) |
| Cursor (code editor) | Standalone IDE (VS Code fork) | Local index, up to 50K files | Mid-scale | Usage-based, ~$40-100/mo | No | 50.21% (Feb 2026) |
| Claude Code | CLI | CLAUDE.md + tool calls | Medium | API usage-based | N/A | 49.75% (Feb 2026) |
| GitHub Copilot | IDE extension | Repo search + recent files | Large | $10-19/mo per user | Yes | Not officially submitted |
| Cline | IDE extension (VS Code) | MCP + file tools, manual context | Small to medium | Open source; API costs vary | No | Not officially submitted |
Cursor's strength is editor-level integration and speed of iteration for prototyping tasks. Claude Code, Anthropic's own CLI agent, excels at complex reasoning tasks and works well for developers who prefer terminal-based workflows. Cline offers maximum model flexibility and transparency as open-source software, with users paying API costs directly. GitHub Copilot has the largest installed base and the lowest friction for adoption.
Augment Code's differentiated position is the combination of large codebase support, enterprise compliance certifications, and the Context Engine's semantic understanding. In independent testing on a 450,000-file monorepo, Augment Code provided deeper cross-service understanding than Cursor or Copilot, though Cursor was faster for rapid prototyping tasks on smaller codebases.
The platform targets organizations where the cost of incorrect AI suggestions is high: large engineering teams working on interconnected microservices, companies with regulatory compliance requirements, and organizations that have already invested heavily in a specific IDE ecosystem they cannot easily replace with a standalone tool like Cursor.
Beyond the core agent, Augment Code has been building out a broader platform for agentic software development.
Cosmos is described as an operating system for agentic software development. It entered public preview in May 2026 for MAX plan users. Cosmos operates across three layers: an application layer with specialized reference agents (deep code review, PR authoring, end-to-end testing, incident response), a system services layer (expert registry, knowledge base, human-in-the-loop checkpoints, learning flywheel, integrations), and a core layer (agent runtime, context engine, event bus, organizational knowledge). The platform allows teams to create custom expert agents, wire them into shared registries, and have them share accumulated knowledge. Unlike a standard agent session that starts fresh each time, Cosmos agents improve through feedback and retain organizational memory.
Intent is a developer workspace launched in April 2026, available initially for macOS on Apple Silicon. It introduces a coordinator-agent model where a Coordinator proposes a living specification before any code is written, fans implementation work out to parallel specialist agents running in isolated environments, and a verifier checks results before returning control to the developer. The intent-to-code workflow aims to reduce the typical eight-step interruption cycle in agentic development to three checkpoints: prioritization, specification review, and deployment approval.
Several limitations have been identified by users and analysts.
The credit-based pricing introduced in October 2025 creates unpredictable costs for high-volume agentic use. A complex agentic task consuming approximately 4,261 credits costs significant fractions of a developer's monthly allocation, and teams running many parallel agents for CI/CD automation can exhaust credits faster than anticipated. Some users reported costs roughly ten times higher after the pricing change compared to the earlier per-message model.
IDE support, while broader than some competitors, is narrower than GitHub Copilot. Augment Code supports VS Code, JetBrains IDEs, Vim/Neovim, and Emacs through ACP, but does not support Visual Studio (the Windows-native Microsoft IDE separate from VS Code), Eclipse, or Xcode natively.
Code completions are being deprecated for most subscription tiers as of March 2026. Developers who rely primarily on inline completions rather than agent workflows may find the feature set narrowing relative to Copilot or Cursor.
Despite strong performance on benchmarks, Augment Code's enterprise customer base is still building public proof points. While the company has named customers across several industries, its publicly documented enterprise deployments are smaller in number than what GitHub Copilot (present across 90% of the Fortune 100) can cite. The SWE-bench Pro advantage over competitors is meaningful but measured in tens of problems out of hundreds, not orders of magnitude.
On extremely large enterprise codebases, performance can degrade at the edges of documented scale, though the company's quantized vector search work was specifically aimed at pushing the practical ceiling higher. Context Engine indexing for very large repositories requires time on first setup, which can be a friction point for initial adoption.
Augment Code's documented use cases span the software development lifecycle:
Feature development: Developers describe a feature in natural language, and Auggie plans implementation, writes code across multiple files, runs tests, and produces a pull-ready changeset. The Quorum case study reported feature timelines shrinking from three weeks to three days on certain customer-driven features.
Codebase onboarding: New engineers on large codebases use Auggie to ask questions about architecture, trace execution paths, and understand why code was written in specific ways using commit history context. One documented case reduced onboarding time on a legacy system from 18 months to 2 weeks.
Legacy system migration: Codem Inc. reported a 50%+ reduction in migration time on a legacy system modernization project. Auggie's ability to maintain context across hundreds of legacy files and apply modern patterns consistently reduces the manual effort of large refactors.
Automated code review: The code review agent runs as a GitHub Action or CI step, posting inline comments on pull requests with codebase-aware feedback. Tekion reported thousands of automated PR reviews monthly.
CI/CD automation: The Auggie CLI in automation mode handles dependency risk assessment, CI failure diagnosis, security audits, and incident response as automated pipeline steps, with MCP integrations pulling in ticket context from Linear or Jira.
Security and compliance workflows: The partnership with Snyk integrates real-time security scanning with Augment Code's autonomous remediation, allowing security issues to be flagged and patched in the same agentic workflow.