Our read
Augment Code suits teams whose hard part is understanding a large repository before touching it: a context engine indexes the code, and changes can start in VS Code, JetBrains, the CLI or Cosmos agents. Review gates stay with your engineers, and the documentation does not detail how dependencies are modeled, so test that on your own code first.
Why the repository comes first
Before an assistant can change code safely, it has to know what that code touches. Pure Storage describes the problem in a vendor-published case study: a monolithic C++, Python and Go repository where, in the words of its CTO for security and GenAI, one small change in a device driver can have large effects elsewhere. Its engineering leaders say Copilot added little value in that repository, and that a cross-organization team chose Augment after evaluating tools. This is the customer's account, not an independent comparison.
Augment's own description puts the context engine at the center: coding agents built on an engine that indexes large codebases. The same engine is named in the Cosmos platform core and in the Auggie CLI, so the understanding step is shared across the entry points rather than rebuilt in each one.
What indexing gives you, and the dependency question
The most concrete example of cross-layer understanding comes from Pure Storage again. Its CTO says other tools could not follow a function wrapped in a C++ wrapper all the way down to a device driver call, and that Augment's context engine could. That is one customer's statement about one codebase.
The pages that describe the context engine give its purpose, not its mechanics. How dependencies are represented, how quickly the index reflects a new commit, and how several repositories are handled are not explained there, so treat them as unknown rather than absent.
Editorial suggestion, not a vendor feature: before committing, pick a function in a shared library used by several services, ask Augment for every caller, and compare the answer with your own search. A miss tells you where a human must keep checking.
Making the change: IDE, CLI or Cosmos
There are three entry points. Extensions exist for Visual Studio Code and JetBrains IDEs, the Auggie CLI brings the agent and context engine into the terminal, and Cosmos runs agents from triggers or schedules. Cosmos ships experts such as a PR Author, which takes a task description and drives it from first commit through merge, and a Tester, which exercises the change and posts results with screenshots. Triggers include GitHub, GitLab, Slack, Linear and Jira, and agents can run in managed sandboxes or on your own machines for code that cannot leave the building.
MoneyGram shows the IDE path. In its vendor case study, an engineer reports using Augment to automate Spring Boot migrations that previously took weeks, now about two hours each, by stating the migration and asking the tool to confirm it is complete. The same account describes Jira and Confluence connections through MCP tools inside IntelliJ and VS Code, and using Augment to explain unfamiliar source instead of tracking down its author. These are one team's reported results, not a benchmark.
The work changes shape: the engineer defines the change and the completion condition, and the tool does the repetitive edits. The documentation read does not describe a step-by-step ticket-to-PR flow beyond the Cosmos experts above.
Where validation happens
Cosmos treats quality gates as part of the design: work pauses for review and resumes on reply, and consequential actions are approved by a person. In the code review loop, every pull request gets a first-pass review so that senior engineers can concentrate on intent and architecture. The vendor attaches a 66% faster time-to-merge figure to that loop without stating the population or baseline, so do not read it as a benchmark.
Code Review is available on all plans and covers GitHub pull requests. Enterprise adds usage analytics, user allowlists, MCP connections to Jira, Linear and Notion for richer context, multi-organization installs and unlimited seats and repositories.
Pure Storage's CTO describes a hobby project in which a system built from formal specifications ran more than 1,000 tests without his reading the code. Note where the trust came from: the specification he wrote and the tests that exercised it, not the tool alone.
What a path costs
Standard is $20 a month and Business is $100 a month, each flat for up to 50 seats, with the same dollar amount included as usage. Usage covers model inference, the Context Engine and Cosmos compute. Model usage is billed at the provider's list price plus a flat 40% service fee, with no fee on compute. Included usage and top-ups are pooled across the team, and top-ups are valid for 12 months. Enterprise is priced by contact.
The consequence is that cost follows activity, not headcount. An agent that reviews every pull request draws from the same balance as engineers asking questions in the IDE, so measure a week of real use before choosing a tier. Standard and Business support is community-based plus a ticket portal; dedicated support is an Enterprise feature.
What stays with the developer
On data, the vendor states that customer code is not used to train models by default and that paid plans exclude AI training under its commercial terms. It lists SOC 2 Type II and ISO 42001, and customer-managed encryption keys and ISO 42001 appear as Enterprise features. Those are vendor statements; request the reports before relying on them.
Everything else is yours: stating the task and what done means, reading the diff, approving consequential actions, and checking that tests exercise the behavior that changed. The product places engineers at checkpoints; it does not decide for them what is correct.
Method and limits
2026-10-28
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.
