Agno Release Notes

Changelog

New Features:

Approvals - Resolved Approval in Post-Hooks: Post-hooks and observability integrations can now read the full resolved approval record (resolved_by, resolved_at, etc.) via run_response.metadata["approval"]. Previously only status and resolution_data were exposed. Lives in metadata so it works uniformly across RunOutput / TeamRunOutput / future WorkflowRunOutput. (#7366, #8032)

Improvements:

PgVector - prefix_match=True actually works: PgVector(prefix_match=True) was a silent no-op - it appended * then routed through websearch_to_tsquery, which ignores wildcards. Now routes through to_tsquery('tok:*') with proper tokenization, so partial queries like "ani" FTS-match "animal" as documented. New cookbook cookbook/07_knowledge/04_advanced/06_prefix_search.py shows the help-center typeahead use case. (#8048, #8051, #8052)

PgVector - empty-tsquery fallback robustness: Replaced the to_tsquery(language, '') fallback with a literal ''::tsquery cast so behavior no longer depends on the parser tolerating empty input. (#8053)

Claude variants - temperature/top_p/top_k = 0 honored: Anthropic, AWS, and VertexAI Claude variants were silently dropping 0.0 values due to bare truthiness checks. Switched to is not None so deterministic output works as intended. (#8009, fixes #8004)

Calendar & Gmail context providers - clean instruction layering: Trimmed DEFAULT_READ_INSTRUCTIONS / DEFAULT_WRITE_INSTRUCTIONS from 157 lines to 27 by removing content already supplied by the toolkit-side instructions. Provider now owns role + safety; toolkit owns the technical reference. (#7982)

decision_log - drop deprecated datetime.utcnow(): Replaced with datetime.now(timezone.utc) to clear the Python 3.12 deprecation warning. (Closes #8030)

Chonkie: Bumped to >=1.6.7 to pick up the upstream language auto-detection fix; removed the temporary test workarounds from #7904. (#8012)

Bug Fixes:

GeminiInteractions - server-side tool calls leaking onto the agent path: On the agent path (Antigravity, Deep Research), FunctionCallSteps describe tools the autonomous loop runs inside the server-managed sandbox. Previously we surfaced them as local tool_calls, leading to Function <name> not found and follow-up 400 invalid_request errors. Gated the FunctionCallStep branch on self.agent is None for both streaming and non-streaming. Model path with user-declared tools is unchanged. (#8045)

Claude (Anthropic / AWS / VertexAI) - temperature=0 silently dropped: See Improvements. Every prior release silently fell back to API defaults (~1.0) when callers explicitly set 0. (#8009, fixes #8004)

What's Changed

• [FIX] Add UTM parameters to links in README by @kyleaton in #8010
• chore: bump chonkie to >=1.6.7 by @sannya-singal in #8012
• cookbook: data_labeling on Gemini + quality review as Workflow by @ashpreetbedi in #8024
• feat: expose resolved approval record to post-hooks via metadata by @brandon-agsys in #7366
• refactor: move resolved approval from typed field to metadata bag by @ysolanky in #8032
• cookbook: image search demo on Gemini + AgentOS by @ashpreetbedi in #8035
• fix: replace deprecated datetime.utcnow() with datetime.now(timezone.utc) in decision_log.py by @Ghraven in #8031
• fix(pgvector): make prefix_match=True actually do prefix matching by @ashpreetbedi in #8048
• [cookbook] Fix Google Calendar spelling by @lil-goat in #8047
• fix: use is not None for temperature/top_p/top_k across all Claude variants by @VANDRANKI in #8009
• cookbook: swap image_search from ChromaDb to PgVector by @ashpreetbedi in #8036
• fix(pgvector): use literal empty tsquery for hybrid_search fallback by @ysolanky in #8053
• cookbook: enable prefix_match=True on image_search PgVector by @ashpreetbedi in #8052
• refactor: simplify Calendar and Gmail context provider instructions by @Mustafa-Esoofally in #7982
• cookbook: add pgvector prefix_match example and improved comments by @Mustafa-Esoofally in #8051
• fix: skip server-side tool calls on GeminiInteractions agent path by @ysolanky in #8045
• chore: Release v2.6.9 by @kausmeows in #8042

New Contributors

• @brandon-agsys made their first contribution in #7366
• @Ghraven made their first contribution in #8031
• @lil-goat made their first contribution in #8047
• @VANDRANKI made their first contribution in #8009

Full Changelog: v2.6.8...v2.6.9

Original source

Changelog

New Features

Antigravity: Added first-party support for Google's Antigravity API in two shapes — AntigravityAgent (a BaseExternalAgent served through AgentOS with native sessions/streaming/UI) and AntigravityTools (a Toolkit that lets any Agno agent delegate a sub-task to a managed Antigravity sandbox).

Gemini Managed Agents: Added GeminiInteractions support for Google's managed agents — Deep Research (autonomous research with citations, background streaming with reconnect/last_event_id resume) and Antigravity (general-purpose agent in a managed Linux sandbox). New agent / agent_config / environment fields, per-agent forcing of background and store, and support for mcp_servers + file_search_store_names on the agent path.

Cookbook - Data Labeling: Added cookbook/data_labeling/ with 18 self-contained workflows covering text, image, audio, video, document, and composed (LLM-as-judge, quality review) labeling primitives.

Cookbook - Slack HITL: Added a deterministic Slack HITL incident-commander demo showing structured pauses via tool_choice="required", user_input echo, and clean termination via stop_after_tool_call=True.

Improvements

Path Safety (Security): Centralized path-safety into a new agno.utils.path_safety module with safe_join and safe_join_subpath. Hardened FileGenerationTools, SlackTools, Toolkit._check_path, agno.skills.utils.is_safe_path, and FileTools.check_escape against path traversal, symlink escape, control-char injection, Windows MagicDot, and Unicode normalization attacks. Introduced PathSecurityError (FileGenerationSecurityError kept as deprecation alias). Bumped requires-python to >=3.9,<4.

Parallel MCP: ParallelMCPBackend now sends User-Agent: agno/<version> on every request so Parallel can attribute traffic to agno.

Evals Config: Dropped the unused available_models field from EvalsDomainConfig; the top-level AgentOSConfig.available_models is now the only supported source for the Evals UI dropdown.

Chonkie: Updated chonkie dependency pin (follow-up to #7869).

Cookbook: Renamed gemini-3-flash-preview to gemini-3.5-flash across the Gemini Interactions cookbooks.

Bug Fixes

AG-UI Teams: Fixed a typo (stream_steps → stream_events) that caused teams running via AG-UI to silently drop all intermediate streaming events (lifecycle, reasoning, tool calls, member delegations) for ~7 months.

Slack Extra: Added aiohttp>=3.7.3,<4 to the slack optional extra — AsyncWebClient imports aiohttp at module load, so pip install agno[slack] was broken for all Slack operations.

Google Drive: Surfaced the Drive API incompleteSearch flag from GoogleDriveTools.search_files() so corpora="allDrives" callers can detect when Drive did not search every drive.

GeminiInteractions: Added a generation_config passthrough field for advanced controls (top_k, presence_penalty, etc.) and fixed history handling — when previous_interaction_id is set, only the new turn is sent rather than replaying the full history. Now actually leverages server-side statefulness.

Anthropic Server Tools: Preserved server_tool_use and code-execution content blocks in message history so multi-turn server-tool flows no longer break across turns.

Anthropic Hardening: Follow-up to the above — switched to the canonical redacted_thinking type literal (the SDK rejects redacted_reasoning_content), accepted both spellings on the streaming start event, added a dedup guard on round-tripped server tool blocks, made coercion failures visible via log_warning, and coerced non-block tool-result items to text.

Gemini Errors: Generic Gemini errors now include the exception type when str(error) is empty or weak, preserving error context across invoke / invoke_stream / ainvoke / ainvoke_stream.

N1N: Added the missing n1n provider mapping in get_model() so Agent(model="n1n:...") model-string paths now resolve correctly.

What's Changed

• [chore] identify agno traffic to Parallel MCP via User-Agent by @NormallyGaussian in #7857
• [fix] Resolve N1N model strings by @pragnyanramtha in #7948
• chore: update chonkie deps by @chonk-lain in #7904
• fix: preserve Anthropic server tool content blocks in message history by @aayushbaluni in #7766
• fix(gemini): preserve generic error context by @MukundaKatta in #7607
• fix: harden Anthropic server tool content block handling by @ysolanky in #7977
• fix: GeminiInteractions generation_config passthrough and stateful input slicing by @ysolanky in #7973
• fix: centralize path safety and harden filesystem-touching tools by @Himanshu040604 in #7707
• cookbook: add deterministic Slack HITL incident commander demo by @Himanshu040604 in #7831
• [fix] Surface incomplete Google Drive searches by @basnijholt in #7824
• fix: add aiohttp to slack extra dependencies by @Mustafa-Esoofally in #7939
• fix: use stream_events instead of stream_steps for AG-UI teams by @Mustafa-Esoofally in #7938
• chore: drop unused available_models field from EvalsDomainConfig by @harshsinha03 in #7993
• cookbook: rename gemini-3-flash-preview to gemini-3.5-flash by @ysolanky in #8005
• cookbook: data labeling primitives (18 workflows) by @ashpreetbedi in #7976
• feat: add Gemini managed agents (Deep Research + Antigravity) to GeminiInteractions by @kausmeows in #7975
• feat: add antigravity tools and external agent by @ysolanky in #8006
• chore: release 2.6.8 by @ysolanky in #8008

New Contributors

@NormallyGaussian made their first contribution in #7857

@pragnyanramtha made their first contribution in #7948

@chonk-lain made their first contribution in #7904

@aayushbaluni made their first contribution in #7766

@MukundaKatta made their first contribution in #7607

Full Changelog: v2.6.7...v2.6.8

Original source

Changelog

New Features

• Gemini Interactions: Added a new GeminiInteractions model class that makes use of Googles new stateful interactions API
• AgentOS: Added an opt-in per-user data isolation layer for AgentOS authenticated endpoints.

Improvements

• Knowledge Readers: Added allowed_hosts parameter to URL-fetching readers to restrict outbound fetches.

Bug Fixes

• Qdrant: Fixed to remove duplicate sparse encoder call in Qdrant async_insert.
• Traces: Fixed the parent trace's session_id / agent_id / team_id from being overwritten by a child agent's spans when both share a trace_id. Most visible when a Team uses a post-hook (e.g. @hook(run_in_background=True).
• Workflow: Fixed workflow’s HITL continue path to use corresponding async function acleanup_run if running in an async context.

What's Changed

• fix: remove duplicate sparse encoder call in Qdrant async_insert by @sannya-singal in #7893
• fix: traces update db level by @kausmeows in #7796
• cookbook: update notion mcp example to current server by @ashpreetbedi in #7921
• feat: per-user data isolation across AgentOS endpoints by @SamJupe in #7606
• fix: Calling acleanup_run instead of cleanup_run by @abhi10691 in #7916
• [fix] fix duplicated word in docker container network_mode comment by @otjdiepluong in #7913
• fix: add allowed_hosts SSRF guard to knowledge readers by @sannya-singal in #7892
• fix: use SESSION_ID_REQUIRED constant in continue-run routes by @ysolanky in #7936
• feat: add GeminiInteractions model for Google's Interactions API by @ysolanky in #7926
• chore: Release v2.6.7 by @kausmeows in #7931

New Contributors

• @otjdiepluong made their first contribution in #7913

Full Changelog: v2.6.6...v2.6.7

Original source

Changelog

New Features:

Slack Interface: Added support for HITL multi-row approvals with all pause types.

Improvements:

WikiContextProvider: Added NotionDatabaseBackend to wiki context provider.

Tools: Updated to warn on duplicate tool names when registering on agent or team.

Bug Fixes:

JWT:

fixed to bind user_id to JWT subject in traces and approvals routers

fixed to bind WebSocket user_id to JWT subject for workflows

RunOutput: fixed aget_last_run_output returns None when agent.id is auto-generated during arun().

/continue Endpoint: Fixed to forward dependencies and metadata to /continue endpoints via get_request_kwargs.

LearningMachine: Fixed to inject LearningMachine context into Team system prompt.

What's Changed

fix: bind user_id to JWT subject in traces and approvals routers by @ysolanky in #7816

fix: bind WebSocket user_id to JWT subject to prevent IDOR by @ysolanky in #7817

feat: add api client header to gemini connectors by @markmcd in #7828

fix: warn on duplicate tool names when registering on agent or team by @ysolanky in #7829

fix: add Anthropic context window patterns to CONTEXT_WINDOW_PATTERNS by @marcospin in #7836

fix: aget_last_run_output returns None when agent.id is auto-generated during arun() by @kausmeows in #7840

fix: forward dependencies and metadata to /continue endpoints via get_request_kwargs by @ysolanky in #7849

fix: inject LearningMachine context into Team system prompt by @Mustafa-Esoofally in #7818

chore: update S3 bucket URL from phidata-public to agno-public by @irfaan101 in #7844

fix: pin tree-sitter-language-pack<1.8.0 to unblock chonkie code chunker by @sannya-singal in #7869

feat: Slack HITL multi-row approvals with all pause types by @Mustafa-Esoofally in #7574

fix: disable agno[mistral] (mistralai quarantined on PyPI) by @harshsinha03 in #7877

[FIX] newsletter link in README by @kyleaton in #7900

cookbook: rewrite 01_demo as minimal AgentOS demo by @ashpreetbedi in #7906

feat: add NotionDatabaseBackend to wiki context provider by @ashpreetbedi in #7914

chore: Release v2.6.6 by @kausmeows in #7915

New Contributors

@markmcd made their first contribution in #7828

@marcospin made their first contribution in #7836

@irfaan101 made their first contribution in #7844

Full Changelog: v2.6.5...v2.6.6

Original source

Changelog

New Features:

Gemini Multimodal File Search: Added multimodal support in the Gemini File Search API (for google-genai≥1.75.0). See cookbook.

Context Providers: Added GmailContextProvider and CalendarContextProvider following the same pattern as existing providers (GDrive, Slack, Database). Also updates GDriveContextProvider to support OAuth in addition to service account auth.

Scheduler: Added scheduler support for Mongo and AsyncMongo Db to run agents, teams, and workflows on a cron in AgentOS.

Workflow Condition Step: Added on_error handling to the Condition workflow step, allowing users to control how errors are handled when sub-steps within a condition fail.

Improvements:

LLMsTxt Tools: Added an allowed_hosts param so agents only fetch from hosts you trust.

SlackContextProvider: Add enable_media_tools flag (default: False) to control file download/upload.

download_file added to read tools

upload_file added to write tools

Bug Fixes:

MCP IDOR: Fixed to bind the user_id on AgentOS MCP tool handlers to the authenticated JWT subject to prevent cross-tenant reads and writes.

Toolkit Instructions: Fixed @tool(instructions=...) to take effect when used inside a Toolkit.

Metrics: Fixed to accumulate parser, output, and member metrics in team run_response.

Workflow: Fixed to persist mid-step workflow run cancellation.

SQLTools: Updated to include column default in describe_table.

Slack Interface: Fixed to replace Slack bot @mention with display name instead of stripping.

GoogleDriveTools: Fixed binary file errors + Office extraction + Shared Drive params in GoogleDriveTools.

ChromaDb: Fixed to offload async batch upsert/insert to worker thread.

Memory: Fixed include identity fields in update_memory and delete_memory tools.

What's Changed

fix: accumulate parser, output, and member metrics in team run_response by @ysolanky in #7704

fix: workflow cancellation persistance by @kausmeows in #7732

fix: include column default in describe_table by @Mustafa-Esoofally in #7703

feat: add search_messages and media tools to SlackContextProvider by @Mustafa-Esoofally in #7702

cookbook: frameworks quickstart for Agno + Claude Code + LangGraph + DSPy by @ashpreetbedi in #7743

fix: replace Slack bot @mention with display name instead of stripping by @Mustafa-Esoofally in #7750

fix: add allowed_hosts allowlist to LLMsTxtTools by @harshsinha03 in #7759

feat: add on_error handling to Condition step by @rotem-bar-cyera in #7214

feat: add Gmail and Calendar context providers by @Mustafa-Esoofally in #7747

fix: binary file errors + Office extraction + Shared Drive params in GoogleDriveTools by @Mustafa-Esoofally in #7764

fix: surface per-tool instructions when registered via Toolkit by @harshsinha03 in #7798

fix: chromadb offload async batch upsert/insert to worker thread by @basnijholt in #7711

fix: include identity fields in update_memory and delete_memory tools by @Mustafa-Esoofally in #6550

[feat] Added scheduler support for mongo by @abhi10691 in #6938

fix: bind MCP tool user_id to JWT subject by @sannya-singal in #7811

feat: gemini file api multimodal by @kausmeows in #7788

chore: Release v2.6.5 by @kausmeows in #7809

New Contributors

@rotem-bar-cyera made their first contribution in #7214

Full Changelog: v2.6.4...v2.6.5

Original source

Changelog

New Features

WorkspaceContextProvider (#7709): Project-aware context provider for repository roots, backed by the read-only Workspace toolkit instead of generic FileTools. Centralizes local filesystem exclude patterns so both FileTools and Workspace skip .context, .venvs, and other agent/dependency/build noise by default. FilesystemContextProvider now accepts exclude_patterns for explicit opt-out/customization.

Improvements

SlackContextProvider (#7710): Simplified provider and inlined SlackTools construction. Removed for_bot_read(), for_assistant_search(), and for_write() factory methods in favor of explicit flags. Added opt-in enable_workspace_search parameter. Tools are now self-documenting via SlackTools, removing runtime agent switching.

Bug Fixes

Slack:

• Gracefully fall back to public channels when the groups:read scope is missing.
• Restore Slack read instruction override and tighten briefing guidance.
• Expose Slack update in agent mode; agent mode returns query-only.

What's Changed

• feat: add workspace context provider by @ashpreetbedi in #7709
• refactor: simplify SlackContextProvider, remove factory methods by @Mustafa-Esoofally in #7710
• refactor: clean up SlackContextProvider and optimize channel resolution by @Mustafa-Esoofally in #7708
• refactor: clean up Slack context provider and add engineering briefing cookbook by @ashpreetbedi in #7706
• Release: v2.6.3 by @ashpreetbedi in #7713
• Full Changelog: v2.6.2...v2.6.3

Original source

Changelog

New Features:

Workspace Tools: Added a polished local-machine toolkit that gives agents read/list/search/write/edit/move/delete/shell access to a root directory tree, with destructive operations gated by Agno's built-in human-in-the-loop confirmation by default.

Improvements:

Models: Updated default id for multiple model providers to newer, actively supported models to prevent upcoming deprecations, improve performance consistency, and significantly reduce operational costs.

Bug Fixes:

Workflow HITL: Fixed to preserve HITL review config in deep_copy for Step, Router, and Loop.

What's Changed

fix: preserve HITL review config in deep_copy for Step, Router, and Loop by @kausmeows in #7694

[feat] migrate to latest provider models and remove deprecated ones by @chiruu12 in #7681

feat: add Workspace toolkit with HITL confirmation gates by @ashpreetbedi in #7683

chore: Release v2.6.2 by @kausmeows in #7701

Full Changelog: v2.6.1...v2.6.2

Original source

Changelog

New Features:

Claude Prompt Caching (Multi-Block): Three Anthropic prompt caching enhancements, scoped to the Claude model (#7662):

New system_prompt_blocks: List[SystemPromptBlock] on Claude — each block has text, cache, and optional ttl ("5m" or "1h") that overrides the model-level extended_cache_time flag per block.

New cache_tools: bool on Claude (Anthropic, AWS Bedrock, VertexAI). Adds cache_control to the last tool so the tool prefix is cached.

Deterministic tool ordering in Model._format_tools (sort by name), so request prefixes stay stable across runs and prompt caches actually hit. Applies across Anthropic, OpenAI, Gemini, and Bedrock.

Parallel MCP Backend: Added ParallelMCPBackend as a new web backend for WebContextProvider (#7667). Talks to Parallel's public MCP server at search.parallel.ai/mcp, exposing web_search + web_fetch (compressed markdown output). Keyless by default; Bearer-auth via PARALLEL_API_KEY for higher rate limits; optional OAuth endpoint via use_oauth=True. Defaults to a 30s timeout (up from MCPTools' 10s) because web_fetch frequently runs longer on large pages.

Improvements:

OpenAI Model String: Mapped "openai:" model string prefix to OpenAIResponses (previously OpenAIChat). Added "openai-chat:" as a fallback for users who still need OpenAIChat. Agent(model="openai:gpt-5.4") now resolves to OpenAIResponses(id="gpt-5.4").

Bug Fixes:

A2A: Pinned a2a-sdk>=0.3.0,<1.0 in the release workflow. a2a-sdk v1.0 has breaking changes that need to be adopted in a follow-up.

Cookbook:

Fixed outdated Redis cookbook docs - removed dependency install and setup commands that are no longer needed (#7615).

Fixed output_schema example to demonstrate output_schema with a Pydantic model (it was incorrectly using output_model with web search).

Fixed typo in cookbook/05_agent_os/README.md that referenced a non-existent agno_agent.py (corrected to agno_assist.py) (#7663).

What's Changed

chore: map "openai:" model string to OpenAIResponses by @kausmeows in #7655

fix typo in 05_agent_os readme by @asifshaikh in #7663

[cookbook] fix outdated docs for Redis by @edlng in #7615

feat: add ParallelMCPBackend for web search via Parallel MCP by @Mustafa-Esoofally in #7667

feat: multi-block prompt caching for Claude by @ysolanky in #7662

chore: Release v2.6.1 by @kausmeows in #7666

New Contributors

@asifshaikh made their first contribution in #7663

@edlng made their first contribution in #7615

Full Changelog: v2.6.0...v2.6.1

Original source

Changelog

New Features:

Team HITL: Added API layer for Team human in the loop and support in AgentOS chat page.

Team Approvals: Added support for Approvals in Team, along with in AgentOS chat page.

Workflow Executor HITL: Added support for executor level human in the loop in case when a pause tool flow is setup on an agent/team in a particular step of a workflow.

Multi-framework support in AgentOS (Beta): Added basic support for ClaudeAgentSDK, Langgraph and DSPy using a single AgentProtocol as the backbone which works with AgentOS runtime.

Reconnection and Resume Runs: Agent/Team running in background using SSE can now be reconnected and resumed in AgentOS in case of a interruption (refresh) so you start from where you left.

Factories: Agent/Team/Workflow can now be created dynamically at run-time using AgentFactory, TeamFactory and WorkflowFactory allowing for more multi-tenant usecases.

Context Provider: Added agno.context — a first-party API for plugging an external source (filesystem, web, SQL database, Slack, Google Drive, MCP server) into anagent as a natural-language tool.

Bug Fixes:

Callable Factory:

Fixed to preserve factory tools and members when Agent/Team is copied.

Fixed to skip iteration of callable-factory tools in Agent.deep_copy.

Telegram Interface:

Fixed to respect retry_after on Telegram 429 rate limit errors

Added quoted_responses option to Telegram interface

Agent OS Router:

Fixed GET /workflows/{id} which omitted nested workflow steps and Condition.else_steps (along with the agents inside them) from the response

Workflow:

Fixed to preserve parent_step_id through nested workflow enrichment layers in case of workflow as a workflow step.

MCP:

Fixed to handle None headers when merging init headers in MCP tools.

Claude:

Fixed to make anthropic imports lazy in agno.utils.models.claude.

Remote Content Source:

Fixed where two uploads of the same filename from different remote sources (GitHub repos, S3 buckets, GCS buckets, etc.) produced identical content_hash values.

FileTools:

Fixed to add exclude_patterns to FileTools.

Agent HITL:

Fixed pause-continue flow not handling subsequent tool calls.

Breaking Changes:

Session Type Filter: The /sessions endpoint now returns all session types (agent, team, and workflow) by default, giving you a complete view of your sessions in a single call. To filter for a specific type, pass?type=agent ?type=team, or ?type=workflow in your request.

Original source

Changelog

Improvements

Option to disable Claude file citations (#7511)

Allow GitHubConfig repo to be specified per request (#7496)

Bug Fixes

Preserve custom db table names when loading components (#7508)

Apply header_provider headers during MCP initialization (#7507)

Preserve inner workflow event identity and add nested_depth to agent/team events (#7491)

Build knowledge dbs live in config API (#7498)

Stop injecting shared HTTP/2 client into all model providers (#7492)

Catch CancelledError explicitly in all router streaming generators (#7379)

Try raw JSON parse before cleaning to preserve code blocks in strings (#7402)

Exclude framework-injected params from user_input_schema (#7485)

Include extra_messages in memory pipeline gate check (#7470)

What's Changed

[fix] Include extra_messages in memory pipeline gate check by @ArielTM in #7470

chore: add weekly scheduled run to main validation workflow by @sannya-singal in #7453

fix: exclude framework-injected params from user_input_schema by @kausmeows in #7485

fix: try raw JSON parse before cleaning to preserve code blocks in strings by @ysolanky in #7402

fix: catch CancelledError explicitly in all router streaming generators by @ysolanky in #7379

fix: stop injecting shared HTTP/2 client into all model providers by @ysolanky in #7492

fix: build knowledge dbs live in config API by @sannya-singal in #7498

fix: preserve inner workflow event identity and add nested_depth to agent/team events by @Ayush0054 in #7491

fix: apply header_provider headers during MCP initialization by @ysolanky in #7507

fix: preserve custom db table names when loading components by @ysolanky in #7508

feat: allow GitHubConfig repo to be specified per request by @ysolanky in #7496

[feat] Option to disable Claude file citations by @joeportela in #7511

chore: release 2.5.17 by @ysolanky in #7494

New Contributors

@ArielTM made their first contribution in #7470

Full Changelog: v2.5.16...v2.5.17

Original source

Changelog

New Features:

LLMsTxtTools and LLMsTxtReader: Adds support for the llms.txt standard - a standardized way for websites to provide LLM-friendly documentation indexes (e.g. https://docs.agno.com/llms.txt).

SalesforceTools: Added Salesforce CRM Tools.

Model: Added new Azure AI Foundry Claude model provider.

OpenAIResponses: Added background mode support for OpenAI Responses API.

Bug Fixes:

Knowledge: Fixed to read knowledge_table from contents_db instead of agent.db.

AGUI:

Fixed to emit reasoning events in AGUI interface.

Fixed AGUI input_content to store current user input instead of message history.

Workflow: Fixed to handle file path images in workflow step conversion.

OpenAIResponses: Fixed to handles response.reasoning_summary_text.delta events to stream reasoning content as it arrives.

Team: Fixed TeamSession.from_dict() to not mutate input mapping.

What's Changed

fix: read knowledge_table from contents_db instead of agent.db by @sannya-singal in #7437

chore: update cookbook to use 5.4 and openai responses model by @kausmeows in #7442

fix: emit reasoning events in AGUI interface by @Mustafa-Esoofally in #7429

[feat] Add Salesforce CRM tools by @raghavenderreddygrudhanti in #7079

[feat] Add Azure AI Foundry Claude model provider by @alejandro-workpath in #6816

[fix] Fix AGUI input_content to store current user input instead of message history by @RowanLane in #5469

fix: handle filepath images in workflow step conversion by @ItsRoy69 in #7457

feat: add background mode support for OpenAI Responses API by @ysolanky in #7444

fix: TeamSession.from_dict() no longer mutates input mapping by @YizukiAme in #7462

fix: stream reasoning summary deltas in OpenAI Responses API by @ysolanky in #7445

fix: sort order enum default in remotedb routes by @Shiv1909 in #7452

feat: add LLMsTxtReader and LLMsTxtTools for llms.txt support by @ashpreetbedi in #7458

chore: Release v2.5.16 by @kausmeows in #7463

New Contributors

@raghavenderreddygrudhanti made their first contribution in #7079

@alejandro-workpath made their first contribution in #6816

@RowanLane made their first contribution in #5469

@YizukiAme made their first contribution in #7462

Full Changelog: v2.5.15...v2.5.16

Original source

Changelog

New Features

Skills: Added skills support to Team. See docs.

Workflow: Added workflow as a workflow step (nested workflow) support. See docs.

Improvements

Workflow HITL:

Workflow steps now support post-execution human-in-the-loop review - pause after a step runs to review, approve, reject with feedback and retry, or edit the output before it continues (requires_output_review on Step, Router, and Loop). See docs.

Parameters consolidated into HumanReview config class - pass human_review=HumanReview(...) on Step, Loop, and Router instead of flat params. Fully backward compatible. See docs

Traceback Logging: Added AGNO_LOG_TRACEBACKS env var (opt-in) to control traceback visibility in log_error and log_warning — off by default (clean logs), full tracebacks when set to true.

SessionSummaryManager: Adds last_n_runs and conversation_limit parameters to
SessionSummaryManager so users can control how much conversation history is included when generating session summaries.

Bug Fixes

OpenAI: Stop injecting shared HTTP/2 client into OpenAI and Azure OpenAI models. Fixes transient 400 errors under concurrent usage by letting the OpenAI SDK manage its own HTTP client.

Metrics: Fixed audio_total_tokens not being computed for OpenAI, Perplexity, and LiteLLM.

Workflow:

Fixed else_steps not being copied for Condition step type.

Fixed StepInput to_dict to properly serialize File objects.

SurrealDb: Updated correct SurrealDB flexible field syntax.

AgentOS: Added missing MIME types for .msg, .xlsx and .xls in AgentOS file upload.

A2A: Fixed to serialize Pydantic output_schema content before str concatenation.

Team: Fixed TeamSession.get_messages(..) returning duplicate messages.

SlackTools: Fixed to add thread-aware messaging instructions to SlackTools.

GitHubTool: Fixed IndexError in get_pull_requests when fewer PRs than requested limit

WikipediaTools: Fixed to handle Wikipedia DisambiguationError and add configurable auto_suggest.

What's Changed

fix: else_steps are not copied for Condition step type by @korntewin in #7198

[cookbook] fix Slack cookbook configs: history context and MemoryManager model by @Himanshu040604 in #7288

[fix] correct SurrealDB flexible field syntax by @pandego in #7303

fix: add missing MIME types for .msg, .xlsx and .xls in AgentOS file upload by @wildchron in #7329

[fix] a2a stream: serialize Pydantic output_schema content before str concatenation by @NIK-TIGER-BILL in #6936

feat: add skills support to Team by @uzaxirr in #7017

[fix] Workflow StepInput to_dict was not serializing File objects by @luisnmartins in #7364

fix: TeamSession.get_messages return duplicated messages by @kausmeows in #7372

fix: add thread-aware messaging instructions to SlackTools by @Mustafa-Esoofally in #7308

Fix IndexError in get_pull_requests when fewer PRs than requested limit by @kaiisfree in #7353

chore: improve exception logging across the SDK by @kausmeows in #7358

fix: remove extra session_type query filter in AsyncMongoDb.get_session by @lethuan127 in #7392

[fix] Add error handling to get_top_hackernews_stories by @lawrence3699 in #7343

fix: stop injecting shared HTTP/2 client into OpenAI-based models by @fehmisener in #7328

[fix] compute audio_total_tokens and correct cookbook metrics examples by @Shiv1909 in #7417

[fix] handle Wikipedia DisambiguationError and add configurable auto_suggest by @Himanshu040604 in #7222

feat: add last_n_runs and conversation_limit to SessionSummaryManager by @ysolanky in #7401

fix: correct stale docstring param name in AgentSession.get_messages by @harshsinha03 in #7423

feat: add post-execution output review and 6 other HITL features for workflows by @ysolanky in #7409

feat: workflow within a workflow by @kausmeows in #6116

refactor: consolidate HITL params into HITL config class by @ysolanky in #7428

refactor: add HumanReview support to Condition, Steps, and Parallel by @ysolanky in #7433

chore: Release v2.5.15 by @kausmeows in #7420

New Contributors

@korntewin made their first contribution in #7198

@wildchron made their first contribution in #7329

@luisnmartins made their first contribution in #7364

@kaiisfree made their first contribution in #7353

@lethuan127 made their first contribution in #7392

@lawrence3699 made their first contribution in #7343

@Shiv1909 made their first contribution in #7417

Full Changelog: v2.5.14...v2.5.15

Original source

Changelog

New Features:

• Fallback Models: Added support for Fallback Models on Agents and Teams. See docs

agent = Agent(
model=OpenAIChat(id="gpt-4o", base_url="http://localhost:1/v1", retries=0),
fallback_models=[Claude(id="claude-sonnet-4-20250514")],
)

• Azure Blob Storage: Added SAS token authentication support for AzureBlobConfig.
• SlackTools: Added workspace search tool to SlackTools.

Bug Fixes:

• LearningMachine: Fixed to filter non-conversational messages from learning extraction.
• OpenAIResponses: Fixed to use correct MIME type for image bytes in OpenAI Responses API.
• Claude: Add automatic trailing user message injection for Claude 4.6+ models that don't support assistant message prefill, with centralized detection across all providers (Anthropic, Bedrock, Vertex AI, LiteLLM).

What's Changed

• feat: add SAS token authentication support for AzureBlobConfig by @ashpreetbedi in #7247
• feat: add fallback model support to Agent and Team by @ysolanky in #7080
• fix: revert IntermediateRunContentEvent changes done in teams by @kausmeows in #7295
• fix: patch review bugs across fallback cookbooks, JWT middleware, and OpenAI error handling by @harshsinha03 in #7299
• feat: add workspace search tool to SlackTools by @Mustafa-Esoofally in #7277
• fix: filter non-conversational messages from learning extraction by @Mustafa-Esoofally in #7280
• fix: use correct MIME type for image bytes in OpenAI Responses API by @Mustafa-Esoofally in #7285
• fix: suppress duplicate member content in Slack team streaming by @Mustafa-Esoofally in #7300
• fix: add text/markdown mime type and fix Anthropic nested schema validation by @ysolanky in #7287
• fix: add inject_trailing_user_message and `trailing_user_message_content for models that reject assistant prefill by @harshsinha03 in #7304
• chore: release 2.5.14 by @ysolanky in #7282

Full Changelog: v2.5.13...v2.5.14

Original source

Changelog

Improvements:

ReliabilityEval: Add subset matching, argument validation, and missing tool call tracking with multi-round tool call collection fixes.

AgentOS: Enhance /sessions list API to return additional fields (user_id, agent_id, team_id, workflow_id, session_summary, metrics, total_tokens, metadata).

AgentOS: Add /info API endpoint to return agent, team, and workflow count as lightweight, unauthenticated instance metadata.

ChromaDB: Implement dynamic batch splitting for large upsert/query operations.

Reader: Propagate chunk_size to default chunking strategies in reader classes.

Slack Interface: Add show_member_tool_calls param and automatic card overflow rotation — rotates to a new message when text exceeds a threshold

Bug Fixes:

VertexAI/Bedrock Claude: Support messages parameter in _prepare_request_kwargs for Claude sub classes

Workflows: Fix continue_run to correctly pause at Condition, Loop, and Router HITL steps — previously only Step instances were checked.

AgentOS: Exclude interface routes (Slack, Telegram, WhatsApp, A2A) from JWT middleware so webhook deliveries are not rejected with 401.

Tracing: Fix trace session stats grouping to use session_id only, preventing duplicate rows when a session has runs from different users.

SurrealDB: Fix trace session stats to use array::first() instead of math::max() for string fields.

ReliabilityEval: Fix multi-round tool call extraction so all rounds are collected; fix mutation bug modifying original RunOutput.messages; fix arun() using wrong ID for file save.

What's Changed

• fix: propagate chunk_size to default chunking strategies in reader classes by @sannya-singal in #7212
• cookbook: offline mode support for Qdrant hybrid search by @sannya-singal in #7181
• fix: implement dynamic batch splitting for ChromaDB operations by @sannya-singal in #7148
• fix: continue_run now pauses at Condition, Loop, and Router HITL steps by @ysolanky in #7227
• fix: exclude interface routes from JWT middleware by @ashpreetbedi in #7252
• cookbook: add HITL config save/load workflow examples by @ysolanky in #7225
• feat: add /info api to return agent, team, workflow count by @kausmeows in #7190
• fix: improve /info endpoint auth bypass and response model by @ysolanky in #7256
• fix: group trace session stats by session_id only by @kausmeows in #7243
• fix: use array::first() for SurrealDB trace session stats by @ysolanky in #7257
• fix: enhance /sessions list API to return additional session fields responses by @uzaxirr in #6270
• feat: add show_member_tool_calls param and card overflow rotation for Slack streaming by @Mustafa-Esoofally in #7244
• feat: add subset matching and argument validation to ReliabilityEval by @harshsinha03 in #7230
• fix: add missing messages parameter to VertexAI and Bedrock _prepare_request_kwargs by @harshsinha03 in #7245
• feat: enable channel summarization in Slack interface by @Mustafa-Esoofally in #7255
• chore: release 2.5.13 by @ysolanky in #7258

Full Changelog: v2.5.12...v2.5.13

Original source