Augment Code review: excellent retrieval, enterprise packaging

Augment has a defensible answer to the hardest part of coding agents: finding the right evidence in a large repository. The purchase case depends on whether that retrieval advantage is worth an enterprise platform, a moving product surface, and usage-based economics.

By the Continuum team. We build a workbench that runs Claude Code, Codex, and their peers, so the model rates quoted here are the ones our own cost analytics ship with.

The short version

Augment Code is strongest as a context and enterprise workflow layer. Its real-time, branch-aware semantic index is technically distinct, Auggie is a capable multi-model terminal agent, and Context Engine MCP can improve agents a team already uses. The weaknesses are packaging churn, a current $100 Business floor, hosted indexing, separate permission systems across surfaces, and a broad product line that can be difficult to evaluate as one coherent seat. Small repositories and individual developers may get better value from Cursor or a bundled Claude Code or Codex subscription. Large teams should run repository-specific retrieval and code-review tests before buying.

What you need to know
  • Retrieval is the strongest reason to evaluate Augment. Hold the model fixed and measure files found, turns, tokens, and corrections.
  • Auggie is competitive as an agent. Its multi-model catalogue, rules, skills, plugins, MCP, and subagents cover serious terminal workflows.
  • The product has changed quickly. Completions were sunset for most plans, Remote Agents are deprecated in analytics, and Cosmos now leads.
  • Current pricing favors teams. Business is $100 monthly for up to 50 seats with $100 of pooled usage.
  • Vendor benchmarks are hypotheses. Augment publishes useful methods and numbers, but buyers need their own repo test.
  • Best fit: large, connected codebases. The value falls when ordinary text search already locates the relevant code.

Verdict by buyer

For a platform team with several services, strict procurement requirements, and agents already wasting time on repository archaeology, Augment belongs on the shortlist. For one developer in a small repository, the current Business plan and platform breadth are hard to justify against a $20 editor or terminal subscription. The product earns its score through retrieval and organization controls rather than a singular editor experience.

Scores reflect the August 2026 product and public evidence, with 10 representing a clear category leader.

AreaScoreReason
Large-codebase retrieval9/10Real-time personal indexes, code-specific embeddings, cross-repo support, and an MCP delivery path
Terminal agent8/10Multi-model Auggie with tools, rules, skills, plugins, permissions, headless mode, and subagents
IDE workflow7/10VS Code and JetBrains support is useful, while non-enterprise inline completions have ended
Pull request review8/10Repository-aware review, custom rules, risk analysis, and remediation paths; independent validation remains limited
Enterprise administration9/10Identity, audit, SIEM, data controls, deployment options, and organization analytics
Individual value5/10The live $100 monthly Business entry point is team-shaped
Pricing clarity6/10The current dollar model is clear, while older plan articles and credit documentation create search-result confusion
Product continuity6/10Rapid moves from completions and Remote Agents toward Intent and Cosmos increase migration risk

What the product does well

It treats context as infrastructure

Many agents search at task time. They list directories, grep strings, open several candidates, and carry those results into later turns. Augment performs semantic indexing continuously and retrieves ranked spans before the model spends turns exploring. That approach is most valuable when a concept crosses naming boundaries, repositories, service layers, and years of history.

The branch-aware design is a practical strength. A global index that knows only main can mislead a developer working on a refactor branch. Augment says it maintains a personal index per developer and updates it within seconds. Its proof-of-possession protocol also requires the client to demonstrate local possession before indexed content can be returned. Those two details address freshness and authorization, which generic retrieval descriptions often omit.

Auggie is a full agent rather than a search demo

The CLI can edit, run commands, call GitHub and other integrations, use MCP servers, run non-interactively, select among several model families, resume sessions, apply project rules, load skills and plugins, and delegate work to custom subagents. It can also serve only the codebase-retrieval tool over MCP. A team can therefore adopt Augment at three depths: retrieval beside an existing agent, Auggie as the agent, or Cosmos as the workflow platform.

The model catalogue reduces one kind of lock-in. Current documentation lists Claude, GPT, Gemini, Kimi, and Prism routing. Augment can switch the model while preserving its Context Engine and tool harness. This makes model evaluation easier than migrating the surrounding workflow every time a lab changes the frontier. It creates a different dependency on Augment's routing, prompts, accounting, and availability.

The enterprise surface is unusually complete

Business and Enterprise include the coding agent, Context Engine, CLI, MCP and native tools, analytics, Cosmos compute, and formal security controls. Enterprise adds custom scale and support. The company advertises SSO, OIDC, SCIM, customer-managed keys, data residency, audit trails, SIEM, VPC, single-tenant, and on-prem deployment options across its current product and security pages. Procurement still needs to verify which controls attach to the exact plan and deployment.

Where the review story is credible

Augment Code Review is more interesting than a diff summarizer because it can retrieve definitions, call sites, patterns, and architecture outside the changed lines. That is the right technical shape for bugs caused by a new implementation violating an old invariant elsewhere in the repository. It installs through the Augment GitHub App, supports custom rules, generates a pull request description, and posts findings into the existing review surface.

The current Cosmos workflow divides review into roles. A risk analyzer estimates whether a pull request can pass with minimal human attention. Deep Code Review searches for correctness defects. Pair Review covers architecture, security, design, and product judgment. A PR Author can respond to findings and repair the change. Shared review memory records feedback so repeated guidance can become repository knowledge.

Review layerUseful forFailure to test
Diff analysisLocal correctness, missing checks, unsafe changed linesNoisy style comments and obvious restatements
Repository retrievalCall sites, invariants, duplicated logic, cross-service impactPlausible but irrelevant retrieved files
Custom rulesTeam conventions and restricted patternsRules that fire without semantic relevance
Risk analysisRouting low-risk and high-risk work differentlyFalse low-risk classification on small dangerous changes
MemoryLearning from accepted and rejected findingsFeedback that overfits one team or becomes stale
Remediation agentTurning a valid finding into a tested patchThe checker and fixer sharing the same blind spot

Augment reports strong benchmark results against Cursor BugBot, CodeRabbit, and other tools, plus internal production results such as three times more output and lower median merge time under Cosmos. The benchmark method and repository set are worth reading. The vendor selected the product, prompts, models, and interpretation, and the production figures come from Augment's own codebase. Treat them as a reason to reproduce the test rather than as a purchasing conclusion.

A fair pilot samples real merged pull requests with known review comments. Remove the human comments, run each reviewer at the commit that was originally reviewed, and label findings blind as correct, useful, harmless, or wrong. Track recall for known defects, precision, duplicate findings, time to first result, remediation quality, and how often a human has to inspect unrelated code to validate the claim.

Weaknesses and risks

The product line moves quickly

In roughly two years Augment has moved from context-aware completions to Agent, cloud Remote Agents, Code Review, Context Engine MCP, Intent, and Cosmos. Several moves are sensible responses to where coding automation is going. The cost is continuity. Completions and Next Edit ended for most plans in March 2026, and current analytics documentation labels Remote Agent metrics as deprecated. Teams should ask which APIs, clients, metrics, and workflows have a supported migration path.

The small-team price is awkward

The live Business plan costs $100 per month whether the workspace has one person or fifty. It includes $100 of pooled usage, so a team can view the platform fee as prepaid consumption. A solo developer still faces a $100 cash floor before comparing value with Cursor Pro, Claude Pro, ChatGPT Plus, or free local tools. Earlier $20, $60, and $200 individual plans still appear in official articles, which makes fresh search results look current after the packaging changed.

Hosted retrieval creates a real trust decision

Proof of Possession is a thoughtful control, and paid plans exclude training. Source code still has to be processed for indexing and inference. Organizations need to inspect data flow, subprocessors, regions, retention, encryption, incident terms, exclusion rules, model routing, and deletion behavior. A security badge cannot answer whether a particular repository is allowed to leave a regulated network.

Surface-specific permissions add operational work

Auggie has granular CLI tool permissions. The documentation says those permissions do not apply to the IDE extension. Slack has external-channel protection and optional allowlists. Cosmos adds its own policy and human checkpoints. A company deploying all of them must maintain a control matrix rather than assume one global setting governs every agent action.

Retrieval can be confidently wrong

Semantic similarity is a ranking signal. It can return an obsolete implementation, a test helper that looks like production code, or a neighboring service with a different contract. Augment says its ranking accounts for active and deprecated code and uses freshness signals. The agent and reviewer still need deterministic verification through types, tests, runtime checks, and human knowledge of the domain.

Support and reliability are part of the product

A hosted index, model router, connector set, and compute platform create several dependencies behind one prompt. During a pilot, record cold indexing time, retrieval latency, model errors, stuck tools, connector failures, and how quickly support resolves a reproducible issue. Business support uses the community plus a ticket portal under the published SLA, while Enterprise adds dedicated support. Trial and beta use receive community support and sit outside that SLA. A team adopting Context Engine as shared infrastructure should price the failure path, status visibility, and vendor escalation process alongside successful task quality.

Recovery quality deserves its own score. Restart the IDE and CLI during active work, interrupt a Cosmos worker, rotate a repository credential, and make a connected source unavailable. Confirm whether the task resumes from durable state, retries safely, or repeats an external action. Agent quality is irrelevant during the hours when orchestration cannot establish which actions already happened.

Pricing and value

As of August 10, 2026, the public Business plan is $100 per month for up to 50 seats and includes $100 of pooled monthly usage across model inference, Context Engine, and Cosmos compute. Model inference is charged at the provider's public API list price plus a 40 percent service fee. Compute is metered separately without that service fee. Usage beyond the included amount can be topped up. Enterprise pricing is custom.

Team shapeBase cash costInterpretation
1 developer$100/moHigh floor unless Context Engine or Cosmos displaces enough separate spend
5 developers$20/person before overageComparable to entry coding subscriptions if pooled usage is enough
20 developers$5/person before overageSeat economics are attractive; usage becomes the main variable
50 developers$2/person before overageAdministration and retrieval can be cheap per seat; active agent usage can exceed the included pool quickly
More than 50Enterprise quoteContract, support, deployment, and volume terms determine value

The included usage should never be divided equally by seats in a forecast. A small number of heavy agent users can consume most of a shared balance while occasional reviewers consume little. Use the billing dashboard as the source of record, because Augment's credit policy says consumption varies by product, workflow, model, configuration, and integrations.

The correct comparison is total workflow cost. Include model and compute overage, engineer time spent searching, review queue delay, administration, migration work when a product changes, and the tools that Augment replaces. A Context Engine MCP pilot can test retrieval without changing the editor or primary agent, which keeps the first evaluation narrow.

Alternatives and combinations

NeedStrong candidateWhy
AI-first editor and predictive typingCursorIntegrated editor, Tab, foreground Agent, background agents, CLI, and web or mobile tasks
Terminal delegation on a Claude subscriptionClaude CodeMature agent loop, permissions, hooks, skills, MCP, subagents, and Anthropic models
OpenAI coding agent across local and cloudCodexCLI, IDE, app and cloud tasks under ChatGPT or API billing
Open-source terminal workflowAider or OpenCodeInspect, extend, and self-host more of the harness
Repository-wide semantic retrieval for existing agentsAugment Context Engine MCPAdopt retrieval while retaining the current model and agent
Vendor-neutral multi-agent operationsContinuumRun several official agents in worktrees with cross-device control and shared usage visibility

Augment combines well with other tools because its retrieval layer is exposed through MCP. A developer can keep Cursor for editing, invoke Claude Code for delegated terminal work, and give both the same Augment search capability. The risk is cost and state duplication: three dashboards, several indexes, overlapping rules, and no automatic proof that every agent saw identical context.

A two-week evaluation plan

01

Select ten representative tasks

Use a mix of cross-service bugs, migrations, unfamiliar ownership, small edits, test failures, and old architecture. Preserve the expected files, tests, and human solution before any agent run.

02

Isolate retrieval

Run the same external agent and model with native search, then with Context Engine MCP. Randomize order and use clean worktrees so the second run cannot inherit the first run's edits or cache.

03

Measure search and outcome

Record relevant files in the first five reads, total tool calls, tokens, time, tests passed, human corrections, and whether the final patch used existing repository patterns.

04

Backtest review

Run Augment Code Review against historical pull requests with known defects. Score findings blind, including noisy comments and missed high-risk issues.

05

Exercise administration

Add and remove users, revoke repository access, change branches, inspect logs, set permissions, test Slack guest behavior, and verify the data deletion process.

06

Forecast from actual usage

Use the live billing dashboard after two weeks. Apply the 40 percent service fee and compute charges, then model the heavy-user tail rather than multiplying an average by headcount.

Buy when the retrieval lift is repeated, review signal survives blind labeling, security controls match policy, and the measured monthly total beats the cost of the current search and review burden. Keep the trial narrow if one of those claims remains unproven.

Questions people ask

It is a strong candidate for large and connected codebases where semantic retrieval, shared context, governance, and review matter. Its value is weaker for one developer in a small repository because the current public Business plan starts at $100 per month.

The Context Engine. It maintains a real-time, branch-aware semantic index and can expose retrieval to Auggie or external MCP agents. That capability can be tested directly by holding the model fixed and comparing task evidence and outcomes.

A high solo price floor, rapid product changes, hosted code processing, separate permission models across surfaces, and reliance on vendor-run benchmarks. Semantic retrieval can also return convincing but irrelevant context, so tests remain necessary.

Enterprise support continued after March 31, 2026. Augment sunset Next Edit and Completions for Indie, Standard, Max, and legacy plans on that date, shifting the broader product toward agents, review, Context Engine, Intent, and Cosmos.

Its repository-aware design is technically credible and Augment publishes benchmark and internal production results. A buyer should still backtest historical pull requests and label findings blind, because the public performance evidence comes from the vendor.

Yes. Auggie can run as an MCP server for Claude Code, Cursor, Codex, and other compatible clients. This is the cleanest way to test Augment retrieval while keeping the existing coding agent and model.

Technically yes. Economically, the August 2026 public plan is optimized for teams: $100 monthly for up to 50 seats with $100 of shared usage. A solo buyer should compare measured retrieval gains with cheaper agent subscriptions.

Test cross-repository retrieval, branch freshness, known pull request defects, permission behavior in every client, model and compute usage, Slack guest handling, access revocation, and the path for exporting or deleting organizational data.

Sources

Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.

  1. Augment Context Engine
  2. Auggie CLI reference
  3. Augment available models
  4. Augment Code Review
  5. Augment pricing
  6. Augment security and privacy
  7. Augment real-time index engineering
  8. Augment feature sunset notice
Try it

Measure the context.
Operate every agent.

Continuum keeps provider agents in separate worktrees with one cross-device review path, so the operating layer stays open while you test which retrieval and model actually win.

free app · your subscriptions · local-first