v0.2.0 — Public Beta

Agents & Streaming

• Agent responses no longer duplicate, fragment, or go missing. Multiple streaming bugs that caused text to repeat, appear in the wrong message, or get lost when messages were queued have all been resolved.
• Agents auto-commit (if enabled) after every turn, not just when a task finishes.
• Agents are locked to the provider they were created with. Switching providers prompts you to start a new agent instead of silently changing models mid-conversation.
• Delegated agents correctly inherit their parent's provider, so child agents no longer get blocked by a mismatch.

Performance

• Large notes no longer freeze the app.

@ Mentions

• Fuzzy search understands file paths. The dropdown is wider, anchored to the input. Clicking a file mention opens it.

File Management

• Create files inline from the sidebar. Delete with undo. Right-click tabs to copy paths or reveal in Finder. "Reveal in Sidebar" works.

Diff Viewer & Changes

• Diff viewer no longer loses your scroll position when an agent edits a file. The Changes tab now shows commit context menus, older commit loading, and proper loading states.
• "What commits are relevant" logic is more robust. You can view previous commits and manually select which ones to see.

Terminal & Shortcuts

• Cmd+backtick / Ctrl+backtick now toggles the terminal even when an input is focused.

UI Polish

• Pasting large blocks of text into chat collapses into an inline chip instead of flooding the input.
• Spaces switcher shows workspace status, repo name, and agent activity.
• Workspace root moved from ~/.workspaces to ~/intent (legacy paths still work).
• Entrance animations on workspace page layout sections.

• BYOA: now includes opencode support (on start page & setting page). this means you can use local models with Intent
• now we render Mermaid diagrams (instead of just our own diagrams)
• migrated specialists agents to a file-based format
• activity log now has live updates, improved styling, and better attribution
• fixed a number of BOYA issue. notably:
  • ModelPicker: Enhanced footer, fallback logic, and settings navigation
  • Settings: Add skeleton loading states for providers and integrations
  • ModelPicker refactor: Make side-effect-free by default with opt-in global updates

and lots of bug fixes/tweaks as usual:

• Prevent diff viewer remount during agent file changes
• Staged files persisting in UI after commit
• Resolve all svelte-check errors across Intent app
• Empty context lines in unified diffs were dropped
• Prevent agent process restart during workspace deletion
• Don't send selected text from chat input as context
• Capture and preserve diffs for activity log events
• Prevent stale branch reuse with proper uniqueness checks
• Fix folder expansion not showing contents after git status sync
• Prevent workspace ID collisions
• Auto-commit agent changes when tasks complete
• Model picker shows actual model name instead of 'Default model'
• Allow sending image-only messages
• Improve ToolCall UI display for file_path and MCP tools
• Specialists now use provider models
• Fix tooltip issues in agent overview
• Add force push option when branches have diverged
• Suppress Monaco editor race condition errors
• Remove icons from Push/Pull buttons in sidebar
• Add clear button to terminal overlays
• Agent overview: Bigger cards, improved tooltips, deduplicated panels

Augment's Context Engine is now available as an MCP server, bringing industry-leading semantic search to every MCP-compatible agent.

In our benchmarks, adding Context Engine improved agent performance by 70%+ across Claude Code, Cursor, and Codex. Whether you use these or any other MCP-compatible agent, you can now give it deep codebase context that makes it write better code, faster, and for fewer tokens.

What's new

• Universal MCP support: Works with Cursor, Claude Code, Zed, Kilo Code, Roo Code, and any MCP-compatible agent.
• Two modes: Local mode for active development (runs Auggie CLI locally). Remote mode for cross-repo context (connects to Augment's hosted MCP).
• 70%+ quality improvement: Measured across 300 Elasticsearch PRs with 900 attempts total.
• Faster and cheaper: MCP-enabled runs consistently required fewer tool calls and conversation turns.

Benchmark results

• Claude Code + Opus 4.5: 80% improvement
• Cursor + Claude Opus 4.5: 71% improvement (completeness +60%, correctness 5x)
• Cursor + Composer-1: 30% improvement

Why it matters

Context architecture matters as much as model choice. A weaker model with great context can outperform a stronger model with poor context. When you give the best models great context, they deliver step-function improvements in production-ready code quality.

Availability

• Available now for all Augment Code users
• 1,000 free requests until the end of February 2026
• Setup takes about 2 minutes

Set it up in 2 minutes →

Enterprise customers needed self-service, programmatic access to their organization’s usage data. Adoption tracking, productivity measurement, cost allocation, and integration with existing analytics infrastructure all required manual workflows, with no API available.

The tenant-based usage analytics API was built to provide this access.

What’s new

• Programmatic access to organization-level usage data
• Available to Enterprise Admins only
• Access enabled through a service account token
• API-based access, not tied to the rest of the product stack

Why it matters

This API makes organization usage data available through a self-service, programmatic interface, where it was previously unavailable. It supports scheduled data access for analytics and reporting workflows that require consistent and stable data retrieval.

GPT-5.2 is now available for all our indie, standard, max and legacy plans

GPT-5.2 is designed for complex reasoning tasks where depth and correctness matter more than speed. It prioritizes careful analysis and consistency across long or multi-step prompts.

Pricing

• GPT-5.2: ~390 credits per task

GPT-5.2 is our highest-cost reasoning model for self-serve users. It performs more extensive reasoning per request and is intentionally slower than lower-cost options.

For simpler or latency-sensitive tasks, we recommend using a lower-cost model.

As workflows grow more complex, a single agent in the CLI becomes a bottleneck. Tasks that could run in parallel — like analyzing code, running tests, and preparing changes — end up serialized, harder to reason about, and slower to complete.

CLI subagents let you split work into focused, concurrent agents — all managed from one CLI session.

What’s new

• Run multiple subagents in parallel from the CLI
• Assign clear, scoped responsibilities per agent
• Automatically share context without manual setup
• Improved isolation, observability, and control

Why it matters

CLI subagents reduce coordination overhead and speed up execution by letting agents do what humans already do: divide work and run it in parallel.

• /about command: View user information and debugging details with a new slash command
• Improved navigation with consistent arrow key and j/k shortcuts across all TUI components

We’ve upgraded the extension’s core state management to better support long-lived conversations and complex agent workflows.

As usage grew, implicit state ownership, hidden lifecycle side effects, and inconsistent async handling began to impact performance and stability. Rather than continuing to patch individual issues, we rebuilt the foundation around a single, transactional state system.

What’s new

• A single source of truth with a typed schema
• Normalized storage for conversations, exchanges, and tool calls
• Transactional updates with batched notifications
• Safer persistence and hydration with explicit lifecycle management
• Selector-based derived state instead of ad-hoc caches

Why it matters

• Long conversations load significantly faster and more reliably
• Complex, tool-heavy threads are more stable, with ~1.2–2× faster execution in some cases
• Fewer edge-case bugs as state evolves over time

Availability

• Available now in the Nightly build
• Not gated; will roll out to stable once Nightly graduates
• Nightly uses a different identifier, so existing local history won’t carry over
• No UI changes yet — this is a performance and reliability upgrade

Today we’re launching Augment Code Review — a high-precision AI reviewer built for large, long-lived codebases. It catches correctness, architectural, and cross-system issues that existing tools miss, while dramatically reducing noise.

Highlights

• Highest accuracy on the only public benchmark for AI-assisted code review — outperforming Cursor Bugbot, CodeRabbit, and others by ~10 F-score points.
• Deep, full-repo context retrieval for both high recall and high precision.
• Prioritizes correctness & architectural issues over style nits.
• Custom rule support to enforce team best practices.
• Adaptive learning based on which comments your developers address.
• Enterprise-ready with far lower noise than typical diff-only bots.

Augment Code Review is available to all Augment Code customers starting today, and free for all paid customers for the next week. It’s also free for open source projects — we believe in supporting the OSS community that powers modern software. To get access, send us a link to your OSS project here.

Install in 3 clicks for GitHub Cloud — read the docs to get started.

Features

• Session Switching: Added `/sessions` command to switch between sessions without restarting CLI
• Shell Configuration: Added `/config` command to configure default shell and startup script
• Keyboard Shortcuts: Added Ctrl+/ for undo and Ctrl+Y for redo in normal edit mode
• Verbose Thinking: Added verbose output mode for thinking summaries to show full agent reasoning
• Terminal Focus Tracking: Added focus detection to hide cursor when terminal window loses focus

Session Management

• Show session ID when closing sessions with command to resume by ID
• Support resuming sessions by ID prefix (unambiguous matches)
• Added `-f` flag to filter session list to current workspace only
• Reversed session list order to show newest sessions first

Improvements

• Fixed text wrapping in tool result summaries
• Truncate web fetch results to 150 characters for better readability
• Increased max length for thinking summary titles to reduce truncation
• Truncate large file views to prevent excessive memory usage

Bug Fixes

• Fixed escape key behavior in ask mode (now properly exits to normal mode)
• Fixed ACP login flow path identification
• Removed duplicate error messages in output
• Suppressed verbose npm output during auto-upgrade failures for cleaner error messages