Deep Agents changelog: what's new each month
Every stable Deep Agents release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 5 months covered; the current month updates daily.
November 2026 · so far
25 releases: 0.2.3 → 0.1.0
November brought substantial improvements to Deep Agents' core functionality, with major updates to the JavaScript runtime, MCP tool handling, and agent communication. The month focused on stability, performance, and better integration with external systems.
- JavaScript state saved by quickjs-rs 0.2.x cannot be restored in 0.3.0; existing threads start with fresh JavaScript context.
- Explicit or saved model selection now required (no longer optional).
- Agent-client-protocol minimum version changed to 1.0.0 in deepagents-acp.
Highlights
- QuickJS runtime upgraded to 0.3.0 with warm interpreter caching and new Node-style fs global for async file operations.
- MCP tool management overhauled with deferred schema loading, trust list improvements, and better approval policy handling.
- Agent communication fixes ensure subagent streaming works correctly and message handling uses LangGraph's delta reducer.
- New chat features include structured clarifying questions, background subagent inspection, and skill-scoped MCP tools.
- Performance improvements across Git attribution, editable-install scanning, and dependency minimum bumps throughout the stack.
New
- Node-style fs global with async readFile, writeFile, readdir, access, and rm methods in JavaScript contexts.
- Unified tool access through single tools.* proxy while maintaining defer_loading support.
- Process-wide LRU cache for QuickJS interpreters to keep them warm between evaluations.
- Structured clarifying questions and background subagent inspection in chat mode.
- Support for disabling per-call timeouts in the code interpreter.
- Shell completion scripts via dcode completion command.
- Agent status reporting to terminals using OSC 7501.
- DeepSeek reasoning levels and media-cpu extra with CPU-only PyTorch on Linux.
Improved
- Compaction input budget now sized to match the selected model.
- Idle compaction settings and offload traces are now filterable and tagged with triggers.
- Git attribution overhead reduced by halving probes, backing off after failures, and scanning the worktree once per evidence probe.
- Editable-install and SDK scans limited to deepagents records for better performance.
- MCP tool counts displayed in /tools headers and marketplace included by default.
- Model refusals now raise errors by default and subagent refusals reported to parent agent.
- OpenAI prompt_cache_key changed from per-thread to shared across threads within a local profile for better reuse.
- Symlink names preserved in virtual filesystem results and symlinks validated against approved search scope.
Fixed
- Streaming of asynchronous subagent runs to apps that join them now works correctly.
- Message-channel handling fixed by using LangGraph's _messages_delta_reducer.
- Uploads containing NUL bytes now stored as base64 in StoreBackend to prevent corruption.
- Interpreter shutdown no longer blocks in CodeInterpreterMiddleware.__del__.
- FIFO files no longer cause FilesystemBackend.download_files to hang.
- Auto mode now resolves relative paths consistently with file tools.
October 2026
31 releases: 0.0.9 → 0.3.11
October brought major expansions to Deep Agents' collaboration capabilities, with new Slack and Discord integrations, scheduled job support, and significant SDK improvements for subagent management. The month also introduced Python 3.14 support and performance optimizations across multiple components.
- deepagents-code now requires Python 3.14 (>=3.14,<3.15); auto-upgrade handles this automatically
- deepagents-talon now requires Python 3.14 (>=3.14,<3.15)
- /summarization-model renamed to /offload model
Highlights
- Added Slack channel adapter with user mentions and sender pairing for DM access
- Introduced scheduled jobs with cron expressions and inline subagent execution
- Expanded async subagent features including status updates and improved parent communication
- Added NVIDIA OpenShell sandbox provider and multiple model recommendations
- Upgraded to Python 3.14 with automatic migration support
New
- Slack channel adapter with outbound user mentions and paired sender access
- Per-chat model switching with /model command
- Scheduled jobs with cron expressions, subagent execution, and channel delivery
- Sandboxed execution via DEEPAGENTS_TALON_SANDBOX environment variable
- Approved one-off ask_for_help consultations
- Batch approvals for concurrent tool calls
- Skill pinning by name in SDK via pinned_skills configuration
- Fork mode support for async subagents
Improved
- Async subagents now report when waiting for humans and show status updates to parent agent
- Subagent concurrency made configurable via CodeInterpreterMiddleware(max_subagent_concurrency)
- Sandbox searches now prefer ripgrep and skip .git directories
- Filesystem permission rules and interrupt_on approvals check both requested and symlink-resolved paths
- Prompt cache expiration notifications and TUI cache-expiring hook support
- Skills load tools only when read, reducing unnecessary overhead
- Async subagent follow-ups can queue without interrupting active work
- Performance improvements to avoid importing langchain libraries at SDK import time
Fixed
- Fixed concurrent writes to local threads with proper locking
- Hook execution now fails closed if context changes when a run resumes
- Corrected skill path inference and skill guide command names
- Fixed text edits to binary files now properly rejected
- Fixed ContextHubBackend.glob to honor the path argument
- Fixed async subagents to pass through webhook configuration properly
September 2026
30 releases: 0.7.12 → 0.7.21
September brought major expansions to Deep Agents' capabilities, with significant additions to conversation management, model support, and reliability improvements across all components.
- Subagent mode renamed from handoff to isolated.
- Task tool calls now reject unknown argument keys instead of silently dropping unexpected input.
- FileInfo.size now correctly represents file size in bytes across backends.
Highlights
- Added Discord channel support with improved gateway failure reporting and new chat-scoped conversation history with configurable storage.
- Introduced subagent conversation forking and flexible subagent execution modes, including background expendable subagents and per-task tool selection.
- Added GPT-6 Astra support and pricing for Fireworks GLM-5.3 models, plus new MCP configuration tools and hot reload support.
- Expanded tool capabilities with native streaming PTC tool calls, improved file handling with better truncation explanations, and new /help and /btw commands.
- Improved workspace isolation, conversation reliability, and MCP safety with better OAuth credential handling and bounded resource operations.
New
- Subagent conversation forking in the SDK and renaming subagent mode from handoff to isolated.
- Discord channel support including gateway failure reporting and chat-scoped conversation history with hybrid search.
- MCP configuration tools, channel-based server authorization, hot reload for MCP configuration, and OAuth/device authentication for Slack and GitHub.
- Background expendable subagents, fresh subagents, per-task tool selection, and on-demand subagent execution.
- A /help command, /btw command for isolated side questions, and currenttime tool with timezone-aware wall-clock cron schedules.
- Inline thread mentions and picker via @@ syntax and targeted conversation deletion support.
- Cache bust timer in the footer and durable Debug Console cost breakdown with token and cost refresh while viewing.
- Support for managing tool approvals through tools.json and persistent LangGraph checkpoints.
Improved
- Renamed subagent mode from handoff to isolated for clarity.
- General-purpose subagents now default to fork mode and Auto classifier defaults by provider.
- Improved workspace isolation by resolving project policy per workspace, scoping project dotenv to workspaces, and isolating LangSmith auth.
- Better workspace configuration with bound conversation support, effective approval mode tracing, and request-time working directory binding.
- Enhanced MCP reliability with hardened OAuth device-flow, credential handling, and bounded configuration updates.
- Improved cost-warning behavior by showing session warnings in a modal and suppressing historical warnings when resuming.
- Made thread browsing and resume flows more resilient with recoverable failures and clearer message count headers.
- Better large tool result previews that clearly explain clipping or truncation, including byte-cap losses and per-line clipping.
Fixed
- Fixed glob sorting to tolerate missing modifiedat values and character counting in ls and glob output.
- Fixed request-time working directory binding and marketplace plugin installs to accept bare relative source paths.
- Fixed invalid partial tool-call patching and updated imputed ToolMessage content in PatchToolCallsMiddleware.
- Fixed local voice transcription and embedding retries, and ensured local model downloads persist.
- Fixed binary content in StateBackend.uploadfiles being tagged as base64.
- Fixed prompt clipboard word deletion behavior and local MCP cache warming to not block the event loop.
August 2026
30 releases: 0.7.2 → 0.7.11
August brought significant stability improvements and new user-facing features across Deep Agents. The month focused on better context management, expanded model support, and improved session handling.
- deepagents-code now requires Python 3.12 or newer.
- deepagents-cli is deprecated in favor of managed-deepagents.
- deepagents-talon now requires Python 3.12 or greater.
- Replaced Gemini 3.6 Flash with Gemini 3.7 Flash in the model switcher.
Highlights
- Added /context-doctor command for inspecting and troubleshooting context consumption issues.
- Introduced Python extension support with discovery, configuration, and runtime hosting capabilities.
- Expanded model support with new options including Meta Muse Spark, DeepSeek V4 variants, and Baseten GLM models.
- Improved session persistence with support for resuming and reconfiguring ACP sessions.
- Enhanced diff readability with word-level highlights and added shell command highlighting in chat input.
New
- Added /context usage report for inspecting context consumption and cache status row for at-a-glance session state.
- Added /context-doctor command for diagnosing and troubleshooting context issues.
- Added /summarization-model command and model picker for configuring the summarization model separately.
- Added tri-state DEEPAGENTSCODEONBOARDING environment variable for code onboarding control.
- Added support for managing config from remote sources and DEEPAGENTSHOME environment variable.
- Added warnings before expensive cold-cache turns and trust user-declared endpoints for cold-cache policies.
- Added resizable chat input by dragging its top border.
- Added trace metadata attribution for skill invocations and editable trace metadata support.
Improved
- Hooks v2 is now generally available with support for loading hooks from installed plugins and multiple lifecycle fixes.
- Auto approval classifier now supports selecting the classifier model and setting a review timeout.
- HITL rejection reasons are now framed for the model with easier reject-with-feedback discovery.
- Price catalog now refreshes hourly in the background with Baseten built-in overrides and local fallback support.
- Improved diff readability with quieter hunks and word-level highlights.
- Made /offload server-owned and interruptible with better progress reporting.
- Improved streaming performance for tool-call arguments to run in linear time.
- Enhanced chat and footer UI behavior with better visibility during rapid typing and improved MCP footer wrapping.
Fixed
- Fixed exact-file delete target resolution in the SDK using first-match-wins behavior.
- Fixed leaked turn coroutines and SQLite handles in session management.
- Fixed Hooks v2 lifecycle issues including session-end teardown bounding and malformed hook resume handling.
- Fixed recognition of in-flight trace messages and forced-compaction forks by preserving session IDs.
- Fixed Auto approval gate path expansion handling and tool argument validation errors are now recoverable.
- Fixed refreshed local context handling after compaction to reduce cache busts.
July 2026
10 releases: 0.1.49 → 0.1.51
July brought significant refinements to Deep Agents, including major architectural changes to simplify agent creation and improve security, alongside UI enhancements for better user experience. The month also saw important updates across all integration partners to align with core library changes.
- TodoListMiddleware no longer included by default in createdeepagent; pass middleware=[TodoListMiddleware()] to restore.
- Default agent prompts now lean with empty base prompt; tool-usage prose duplicating schemas removed; BASEAGENTPROMPT deprecated (removal in 0.9.0).
- Built-in tool-usage prompt constants TASKSYSTEMPROMPT, ASYNCTASKSYSTEMPROMPT, SUMMARIZATIONSYSTEMPROMPT, FILESYSTEMSYSTEMPROMPT, and EXECUTIONSYSTEMPROMPT removed.
- Removed deprecated backend compatibility shims; callers must pass concrete BackendProtocol instances and configure StoreBackend with explicit namespace.
Highlights
- Major version 0.7.0 overhaul: removed default todo middleware, simplified default prompts, and improved filesystem security with virtual mode and destructive delete support.
- Added support for LangSmith Gateway credentials and expanded Hooks v2 with client and server lifecycle events.
- Improved user interface with long message handling, better toast notifications, clickable thread IDs, and session cost display in status bar.
- Enhanced filesystem capabilities: writefile now creates missing files and replaces existing ones without errors.
- Better goal handling that resolves underspecified objectives from conversation context.
New
- Slash commands for mode changes and escape option to manual mode from YOLO notice.
- Hooks v2 with workspace trust, client and server lifecycle events, and runtime feedback.
- Thread message counts visible in Debug Console.
- Direct display of askuser answers on the answered tool row.
- Backward navigation: after switching threads, Deep Agents points back to the previous thread.
- Install subcommand added.
- Option to mute the YOLO active toast.
- Claude Opus 5 model support.
Improved
- Long transcript user messages now handled with toast notifications and expansion options.
- Model no-op notices shown as toasts instead of inline messages.
- Server graph bootstrap calls moved off the event loop for better performance.
- Splash screen thread ID is now clickable to copy.
- Filesystem paths in LocalShellBackend and FilesystemBackend anchored under rootdir with traversal protection.
- Goal handling improved to resolve underspecified objectives from conversation context.
- Hook resume stability enhanced across identity and Command tool results.
- Empty read now returned for degenerate readfile windows.
Fixed
- Toast notification shown when submitting empty required askuser answer.
- Workspace trust and Hooks v2 stability improvements across identity and Command tool results.
- Server hook state kept out of task results.
- Post-install restart modal responsiveness maintained.
- Goal criteria prompt kept responsive.
- Turn released when an interrupted worker never starts.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Deep Agents: pricing, features and alternatives · All changelogs