Augment Code

Before executing any code changes, the agent analyzes the entire codebase context (4,456 sources filtered to 682 relevant via semantic understanding) and generates a sequential task decomposition plan (e.g., 5-step OAuth flow: analyze auth → create handler → update middleware → add rotation → write tests). The plan is presented to the user for review, modification, or approval before implementation begins. This prevents blind implementation and allows users to redirect the agent mid-task at any checkpoint.

Maintains a live, semantic understanding of the entire codebase including code dependencies, architecture patterns, documentation, coding style, and recent changes. Processes 4,456 sources and filters to 682 relevant files using semantic understanding (mechanism unspecified — likely vector embeddings or AST-based analysis). Surfaces memories (learned patterns, conventions, past decisions) before saving, allowing users to approve, edit, or discard them. Approved memories become workspace 'Rules' shareable with the team, preventing outdated patterns from persisting across sessions.

Provides SOC 2 Type II compliance (all plans), ISO 42001 compliance (Enterprise), CMEK (Customer-Managed Encryption Keys) for data at rest, SIEM integration, data residency options, granular access controls, comprehensive audit trails, and enterprise SSO (OIDC, SCIM). All plans include 'No AI training allowed' guarantee, preventing customer code from being used to train models.

Assists with architecture-level changes and design reviews, not just file-level edits. Claimed capability to handle complex engineering tasks including architecture and debugging. Example shown: JWT refresh token rotation (multi-file, cross-cutting concern). Design review mode shown in Intent UI example, suggesting capability to analyze and suggest architectural improvements.

Assists with bug identification, root cause analysis, and fix implementation by leveraging semantic codebase understanding. Claimed as core capability ('complex engineering tasks including architecture and debugging'). Integrates with terminal execution to run tests, linters, and debugging tools. Checkpoints allow iterative debugging with reversible changes.

Generates and modifies code across multiple files in a single task while maintaining semantic consistency (e.g., updating auth.ts, session.ts, and middleware in one OAuth flow implementation). Changes are staged at checkpoints after each step, allowing users to accept, revert, or redirect the agent without losing prior work. Implementation phase between checkpoints runs without interruption, but no changes are committed until user approval at each checkpoint.

Executes shell commands and invokes external tools (e.g., build systems, linters, test runners) as part of task implementation. Tool invocation is supported via MCP (Model Context Protocol) and native tool bindings (unspecified which tools are natively supported). Commands are visible in the implementation phase UI and can be reviewed before execution. Sandboxing and execution environment isolation are undocumented.

Analyzes pull requests and generates code review feedback including PR summaries, inline comments, and suggestions for improvement. Operates in two modes: auto mode (generates review without user intervention) and manual mode (user reviews and approves before posting). Review guidelines can be customized per workspace. Integrates with GitHub for multi-org PR operations and supports Slack notifications.

Saves task progress at each checkpoint, allowing users to close the IDE or CLI and resume work the next day with full context recovery. Workspace isolation ensures that each Intent (task/project) maintains separate context, memory, and Rules. Users can 'come back tomorrow and pick up exactly where you left off' with the agent understanding prior decisions and context without re-explanation.

Allows users to select between Claude Opus 4.6 (primary, highest capability) and Gemini 3.1 Pro (marketed as 'frontier AI at half the cost'). Model selection appears configurable per task or workspace (UI shows selectable dropdown). Pricing is credit-based (40,000–450,000 credits/month depending on plan) with auto top-up at $15 per 24,000 credits. Credit-to-token conversion and per-task cost are opaque.

Deploys as VS Code extension, JetBrains IDE plugin, or CLI tool, all using the same underlying Context Engine and agent logic. Users can switch between IDE and CLI without losing context or workspace state. CLI deployment enables terminal-based workflows and CI/CD integration (unspecified).

Integrates with GitHub to perform PR operations (review, analysis, merge suggestions) across multiple organizations without re-authentication. Slack integration enables notifications and command-based task triggering. GitHub integration supports PR summaries, inline comments, and multi-org repository access.

Achieves 51.80% on SWE-Bench Pro leaderboard (highest among agents on Claude Opus 4.5), demonstrating production-ready code generation on real-world GitHub issues. Blind study on Elasticsearch (3.6M LOC, 2,187 contributors) validates performance on large, complex codebases. However, metrics show variance: correctness -11.8%, completeness -11.8%, code reuse -12.4%, best practice -16.4% vs. human baseline.