February 2026 was a landmark month for Agno. We hit 400 contributors on GitHub, released the transformative v2.5.0 with Team Modes and HITL for Teams, changed our license to Apache 2.0, and watched the community ship an incredible wave of production-ready applications.
Whether you're orchestrating multi-agent teams with different execution modes, implementing human-approval workflows, or scheduling autonomous agent runs with cron, February delivered the production infrastructure you need to build real AI systems.
The community responded by shipping everything from autonomous AI marketplaces to voice-enabled RAG systems to multi-paper research assistants.
New releases & features
v2.4.4 - v2.5.4
February delivered ten major releases, including our landmark v2.5.0. Here's what we shipped:
Workflows HITL (v2.5.4)
Add support for HITL in workflows at Step level to pause for confirmation and user input:
- Step-level pausing: Pause workflows at any step for human approval
- Confirmation flows: Require human confirmation before critical workflow steps
- User input collection: Gather information from users mid-workflow
- Seamless resumption: Continue workflow execution after human interaction
Essential for workflows that require human oversight at specific decision points.
Metrics System Redesign (v2.5.4)
Redesigned metrics system to provide per-model, per-component granular tracking across the entire agent/team/workflow lifecycle:
- Granular tracking: Per-model and per-component metrics
- Lifecycle coverage: Track metrics across agents, teams, and workflows
- RunMetrics replacement: Replaced heavyweight RunOutput collectors with efficient RunMetrics
- Team metrics: Include member response metrics in team session metrics
Better observability and cost tracking for production deployments.
TeamMode.tasks Streaming (v2.5.4)
Added streaming event support for TeamMode.tasks, enabling real-time event emission during autonomous task execution. Watch tasks get assigned and completed in real-time.
PgVector Similarity Threshold (v2.5.4)
Added similarity_threshold parameter for filtering search results by minimum similarity score. Only return results that meet your quality bar.
DuckDuckGoTools Enhancements (v2.5.4)
Exposed timelimit, region, and backend params in DuckDuckGoTools for more control over web searches.
SeltzTools SDK Update (v2.5.4)
Updated to use Seltz SDK 0.1.3 for improved search capabilities.
Team Execution Modes (v2.5.0)
Introduced TeamMode enum with four distinct execution patterns:
- coordinate (default): Supervisor pattern where the leader picks members, crafts tasks, and synthesizes responses
- route: Router pattern where the leader routes to a specialist and returns their response directly
- broadcast: Broadcast pattern where the leader delegates the same task to all members simultaneously
- tasks: Autonomous task-based execution where the leader decomposes goals into a shared task list
This transforms how you build multi-agent systems. Different problems need different orchestration patterns—now you can choose the right one.
Human-in-the-Loop (HITL) for Teams (v2.5.0)
The most-requested feature: comprehensive HITL support for teams with improved RunRequirement system:
- Tool confirmation flows: Pause for approval before critical operations
- User input collection: Gather information with schema validation
- External tool execution: Delegate to human-executed tools
- Session state preservation: Fixed HITL pause handlers silently dropping session_state changes
Essential for production teams that require human oversight before high-stakes decisions.
Approval System (v2.5.0)
New @approval decorator for systematic human approval workflows:
- @approval(type="required"): Blocking approval that pauses the run until resolved via approvals API
- @approval(type="audit"): Non-blocking audit trail for compliance/logging
- Persistent approval records with status tracking (pending, approved, rejected, expired, cancelled)
- Approvals API for listing, inspecting, and resolving approvals
Perfect for financial services, healthcare, and other regulated industries that need audit trails.
Scheduler (v2.5.0)
Cron-based scheduling for agents, teams, and workflows:
- Cron expressions for flexible scheduling
- Retry logic for failed runs
- Timeout support for long-running tasks
- Timezone handling for global deployments
Run your agents on a schedule, not just on demand.
LearningMachine for Teams (v2.5.0)
Extended LearningMachine support to team-based agents for persistent learning across team interactions. Teams can now learn and improve from every conversation, just as individual agents do.
Callable Factories for Tools, Knowledge, and Members (v2.5.0)
Tools, knowledge, and team members can now be defined as callable factories resolved at runtime with access to RunContext:
- Runtime resolution: Configure tools/knowledge/members based on session context
- Caching support via
cache_callables - Custom cache keys for session-scoped resolution
Enables dynamic agent configuration based on user permissions, session state, or runtime conditions.
Knowledge Isolation (v2.5.0)
Added isolate_vector_search flag to Knowledge class:
- When enabled, documents are tagged with
linked_tometadata during insert - Searches filter by this tag automatically
- Multiple Knowledge instances can share the same vector database while maintaining isolated search results
- Defaults to False for backwards compatibility
Critical for multi-tenant deployments where knowledge bases must be strictly separated.
License Change to Apache 2.0 (v2.5.2)
Changed license from Mozilla Public License 2.0 (MPL 2.0) to Apache Software License 2.0. This makes Agno more permissive for commercial use and aligns with industry-standard open source licensing. A major milestone for production adoption.
CodingTools (v2.5.1)
New toolkit for code-related operations. Build agents that write, analyze, and refactor code.
UserFeedbackTools (v2.5.1)
New toolkit for collecting user feedback. Essential for building agents that learn from user interactions.
SeltzTools (v2.4.5)
Integration with Seltz Search for advanced search capabilities.
UnsplashTools (v2.4.4)
Search and retrieve high-quality, royalty-free images from Unsplash API. Perfect for content generation agents.
Workflow Condition Step Else Path (v2.4.7)
Added support for an Else path in Condition steps. Define alternative paths instead of just skipping execution.
CEL Support for Workflow Steps (v2.4.8)
Use CEL (Common Expression Language) expressions as evaluators in Condition, Loop, and Router steps. CEL expressions are strings, making workflow steps fully serializable.
Workflow Router Step Enhancements (v2.4.8)
Added the ability to return the name of the Step instead of the step itself, plus support to return a group of steps as a choice for the router.
Neosantara Model Provider (v2.4.8)
Added support for Neosantara, an Indonesian LLM gateway that provides an OpenAI-compatible API. Expanding global reach.
Moonshot Model Provider (v2.4.4)
Added moonshot.ai model provider for additional model options.
AWS Bedrock Reranker (v2.4.7)
New AwsBedrockReranker class supporting Cohere Rerank 3.5 and Amazon Rerank 1.0.
AWS Bedrock Embedder Updates (v2.4.7)
Updated AWS Bedrock to support Cohere v4 Embed.
Remote Knowledge (v2.5.3)
Added endpoints to support the listing of remote contents via API. Enables uploading of content in S3 buckets via AgentOS.
PDF Reader Improvements (v2.5.3)
Added sanitize_content parameter (enabled by default) that normalizes fragmented PDF text extraction—collapses word-per-line artifacts while preserving paragraph breaks. Disable sanitize_content=False for PDFs containing structured content, such as code or tables.
WebsiteReader Chunking (v2.4.8)
Now uses FixedSizeChunking instead of Semantic by default, removing the need for an OpenAI API key when not setting a specific chunker.
Studio (v2.4.8)
Visual editor for building Agents, Teams, and Workflows in AgentOS. Drag-and-drop interface with a Registry for managing tools, models, databases, and schemas.
AWS EFS Support (v2.5.0)
Adds support for AWS EFS volume and mount point for persistent storage in AWS deployments.
MCP Tools HITL (v2.4.7)
MCPTools now works with requires_confirmation_tools. Human-in-the-loop for external tool execution.
External Execution Silent Parameter (v2.4.4)
Added a new parameter to make external tool execution silent by not populating the run response content with placeholder strings.
Community projects & showcases
In February, there was an explosion of production-ready applications. Here are the standout builds:
1. paper2saas by Agentra Labs
AI agents that dissect papers, brainstorm killer SaaS ideas, validate markets, crush competitors, and spit out MVP roadmaps—all hallucination-free:
The workflow:
- Takes arXiv papers as input
- Analyzes research for commercial potential
- Generates SaaS ideas grounded in the research
- Validates market opportunities
- Produces competitive analysis
- Outputs actionable MVP roadmaps
Why it matters: Bridges the gap between academic research and commercial applications. Turns theoretical breakthroughs into business opportunities.
2. OpenClaw Replacement by bitdoze
Built a lightweight AI Assistant alternative to OpenClaw using Agno & Discord:
The setup:
- Host: $5 VPS (Ubuntu)
- Interface: Discord (via Bot API)
- Tools: Web search, Shell access, GitHub, File system
- Automation: Hourly tech/AI news updates via Cron
Why it's cool: OpenClaw is powerful but complex. This proves you can build production-ready AI assistants with Agno on minimal infrastructure—fast, 100% customizable, and budget-friendly.
Tech stack: Agno, Discord Bot API, UV for package management
Full Tutorial | Video Walkthrough
3. OpenAgent by Adam (ajshedivy)
Dedicated fork of OpenCode with AgentOS integrations called openagent:
Features:
- Bring AgentOS to your terminal with
/agnohub - Browse agents with keyboard navigation
- Quick-connect with status indicators
- Agent detail views
Roadmap: Teams, Workflows, and eventually evals/metrics/knowledge
Bonus: Building agentos-sdk to improve the dev experience
Why it matters: Terminal-native AgentOS experience. Developers can manage their AI agents without leaving the command line.
4. agno-client by rodrigocoliveira
Simplified React SDK for Agent OS — solving the "how do I build a frontend for my agents" problem:
The problem: AGUI's CopilotKit integration is complex, and Agent-UI hasn't been updated for full Agent OS capabilities.
The solution: Clean React packages for polished interfaces connected to Agno agents:
- Simple React hooks and components
- Clean abstractions over underlying complexity
- Drop-in solution for beautiful agent UI
- Fully customizable components
- Always in sync with AgentOS FastAPI endpoints
Why it matters: Frontend development for AI agents shouldn't be harder than building the agents themselves. This makes it trivial.
5. Elite Research Agent by Build Fast with AI
Publication-quality research from a single prompt:
How it works:
- Searches the web for relevant sources
- Reads full articles (not just snippets)
- Produces NYT-style investigative reports
Tech stack: Kimi K2.5 (via OpenRouter), Agno, web search tools
Why it's impressive: Demonstrates high-quality reasoning models integrated with Agno for long-form content generation.
6. Production Agent Showcase by Build Fast with AI
Multiple production-ready agents demonstrating different models and use cases:
StepFun Step-3.5-Flash Integration
Built and tested with StepFun Step-3.5-Flash (free via OpenRouter). Full notebook showing:
- Tool calling
- Memory management
- Multi-agent teams
- Knowledge bases
- Structured outputs
Qwen3-Coder-Next Agents
80B MoE model with only ~3B active params → production-grade coding agents without massive API bills.
Tic-Tac-Toe AI Arena
Claude Opus 4.5 vs Kimi 2.5 — not just a game, a benchmark for agentic reasoning:
- Tech stack: Agno, Streamlit + custom Neon CSS, OpenRouter
- Purpose: Test AI reasoning in real-time gameplay
HackerNews Digest Agent
Before: 20 mins scrolling HackerNews daily. After: 10-second AI digest.
The agent:
- Fetches trending stories
- Summarizes each in 2-3 lines
- Groups by theme (AI, DevTools, etc.)
- Explains why you should care
Tech stack: Kimi GLM-5, Agno
7. AgentBazaar by Harish Kotra
Fully autonomous AI Marketplace running locally—agents hiring each other:
The workflow:
- Broker Agent: Takes the request and creates the technical spec
- Bidding War: Multiple Worker Agents (different personalities) generate bids
- Negotiation: Negotiator agent bargains to lower prices
- Escrow & Validation: Validator checks work against the contract before releasing funds
Why Agno? Enforcing output_schema using Pydantic models meant agents weren't just chatting—they were exchanging valid, executable JSON objects.
The result: Self-contained economy with reputation systems, bad-actor filtering, and self-optimizing prices. All running locally on llama3.2 with Ollama.
Vision: Prototype for how future software systems might self-assemble. Instead of monolithic code, imagine micro-services that negotiate their own APIs and SLAs dynamically.
8. Stock AI Agent by Jagdeep Singh
Web-enabled stock analysis agent that fetches real-time data and structured insights:
What it does:
- Searches the web for stock data
- Fetches current prices
- Retrieves analyst recommendations
- Returns structured insights in tabular format
Tech stack: Agno, DuckDuckGo web search, YFinanceTools, Groq, OpenAI
Key learning: Coming from ML model background, this project helped understand how agents work in real-world application workflows, not just notebooks.
9. AI Stock Screener & Interview Agent by MarcelloDOP
Two production applications built with Agno:
Stock Screener Agent
Features:
- Uses YFinance and Tavily for data
- Suggests the best stocks based on fundamental analysis
- Analyzes analyst ratings and news sentiment
Competency-Based Interview Agent
Helps job seekers prepare for interviews:
- Generates personalized questions from the job description and CV
- Creates tailored answers
- Exports to Word document format
Tech stack: Agno, OpenAI (gpt-5-mini), ChromaDB, SQLite, Streamlit, python-docx
10. LLMOps Architecture by Kameshwara Pavan Kumar Mantha
Production-ready GenAI stack showing what enterprise LLMOps actually looks like:
Key technologies:
- MLflow: Unified LLMOps backbone (AI Gateway routing, agent serving, prompt registry, evaluation, observability, tracing, telemetry, cost tracking)
- Agno: Agent framework
- Multi-LLM: OpenAI, Anthropic, Gemini with traffic splitting and fallback
- Qdrant: Vector database for RAG
What it enables:
- Clear visibility into every response
- Structured evaluation of agent quality
- Controlled multi-model routing
- Fully self-hosted stack (free to run)
Why it matters: Practical blueprint for building production GenAI systems with confidence and control.
11. Math Mentor AI by Manoj Kumar Karumanchi
Multimodal RAG + Multi-Agent Math System—not a chatbot, a structured reasoning system:
Architecture:
- Multimodal input: Text, OCR (EasyOCR), Audio (Whisper)
- Agent pipeline: Extract → Guardrail → Parser → Router → Solver → Verifier → Explainer
- RAG stack: Hugging Face embeddings + LanceDB
- Tool execution: SymPy for deterministic math solving
- Model abstraction: Groq + Gemini fallback
- Memory loop: SQLite + embedding similarity retrieval
- HITL checkpoints for ambiguity and low confidence
Why it's interesting: The pattern maps directly to real-world systems like OCR-based document processing, compliance validation, multimodal support automation, and human-in-the-loop AI governance.
Tech stack: Agno, Python, Streamlit, Docker, Hugging Face Spaces
12. Voice-RAG by Shivam Singh
Turn any document into a real-time, source-grounded voice assistant:
What it does:
- Upload a document
- Auto-index it for retrieval
- Provision a voice-enabled agent
- Open a live speech-to-speech widget
- Have a real-time conversation with your knowledge base
Key features:
- Extended Agno plugin with Voice + RAG orchestration
- Integrated ElevenLabs real-time conversational voice (200ms latency)
- Automated ingestion → indexing → agent provisioning
- Embeddable widget
- Multilingual embeddings
Use cases:
- Enterprise support (faster resolution, fewer tickets)
- Education (ask instead of search)
- Healthcare (voice access to clinical guidelines)
- Field operations (instant access to SOPs)
- Research (upload papers, ask follow-ups)
Tech stack: Agno, ElevenLabs Conversational AI
13. PaperSphere-AI by Saathvik Krishnan
Agentic multi-paper RAG pipeline for research understanding and comparison:
The problem: Most paper assistants can answer single questions but struggle to compare multiple papers on results, datasets, and limitations.
The solution:
- Upload one or more research PDFs
- Hybrid retrieval using BM25 + vector search
- Reranking and chunk filtering for better context quality
- Single-paper grounded Q&A with citations
- Multi-paper comparative synthesis with metric-aware outputs
- Evaluation signals for retrieval quality and groundedness
Tech stack: FastAPI (Render), Streamlit (Community Cloud), Qdrant Cloud, Gemini, Agno
Key focus: Reducing hallucinations, improving traceability through citations, and making multi-paper comparison useful for actual research work.
14. Full-Stack AI Chat App by Shubham Vedi
Production-ready AI agent stack with Agno Skills:
Features:
- Agno Agent Skills (modular, hot-reloadable)
- Claude Sonnet 4.6 via OpenRouter
- Tavily for real-time web search
- Streaming responses (SSE)
- Neubrutalism UI design
- One-command startup
- Markdown rendering
Tech stack: Agno Skills, Claude Sonnet 4.6, OpenRouter, Tavily, React, FastAPI
15. FrankenAgent Lab by Ilkan Yıldırım
Modular AI agent builder where you assemble agents like high-tech creations:
The metaphor:
- 🧠 Head: LLM brain
- 🦾 Arms: Tools & integrations
- 🦿 Legs: Workflows & execution
- ❤️ Heart: Memory
- 🦴 Spine: Guardrails & safety
What you can do:
- Build agents visually using YAML blueprints
- Run them via CLI or API
- Trace every tool call
- Share blueprints in the marketplace
- Deploy to production with Cloud Run + Cloud SQL + Redis
Tech stack: FastAPI, Agno, Web UI, execution tracing, user auth, API keys
Why it's cool: Structured way to experiment with agent architectures with production deployment built in.
16. Elite Financial Intelligence Coach by Mihir Kapile
An AI-native platform that transforms raw transaction ledgers into a governed financial roadmap:
Key features:
- Agentic Orchestration: Powered by Agno and Llama 3.1
- Sub-Second Inference: Leveraging Groq's LPU architecture
- Fiscal Stress Test: Simulates capital depletion events
- Zero-Trust Approach: In-memory processing, fuzzy header mapping, encrypted-ready PDFs
Security focus:
- In-memory processing (no database persistence)
- Anonymizes banking formats locally
- User maintains total control over financial data
Tech stack: Groq (Llama-3.1-8B-Instant), Agno, Streamlit, Pandas
Submission for: Palo Alto Networks Case Study
17. OCR Translation System by Qubrid AI
Production-grade OCR-powered language translation system:
What it does:
- Extract text from images using Hunyuan OCR 1B
- Automatically detect source language
- Translate into 48+ languages
- All without GPU setup or infrastructure management
Architecture:
- 4-layer production structure (UI → Pipeline → Agents → Models)
- Multimodal OCR + language models
- Streaming responses for real-time UX
Tech stack: Qubrid AI serverless APIs, Agno, Hunyuan OCR 1B, GPT-OSS-20B
Contributor shoutouts
February saw continued growth with contributions from our expanding community. We hit 400 contributors on GitHub — a massive milestone!
Top community code contributors
- @rodrigocoliveira (GitHub) — agno-client React SDK, HITL session state fix, silent external execution, Firestore session fixes
- @ajshedivy (GitHub) — MCP HITL functionality, OpenAgent terminal interface
- @theagenticguy (GitHub) — AWS Bedrock Reranker and Embeddings expansion
- @WilliamEspegren (GitHub) — Seltz tools integration
- @dvy246 (GitHub) — Unsplash image search tools
- @Tuyohai (GitHub) — Moonshot model provider
- @Byunk (GitHub) — Firestore session storage fixes
- @salimhamed (GitHub) — Workflow run_input JSON serialization
- @tanmaydarmorha (GitHub) — AzureOpenAI authentication enhancements
- @Samyak2 (GitHub) — WebsiteReader OpenAI requirement removal
Feature implementers
- @mklrlnd424 (GitHub) — TavilyTools deprecated parameter removal
- @themavik (GitHub) — Knowledge filename pattern matching, reasoning content deduplication
- @ArivunidhiA (GitHub) — Anthropic streaming metrics fix
- @guguoyi (GitHub) — Cookbook typo refinements
- @jss367 (GitHub) — Documentation link fixes
- @kepler (GitHub) — Session state extraction from run_context, LearningMachine Claude/Gemini compatibility
- @Programocles (dsarlo) (GitHub) — AWS Bedrock Guardrail support
First-time contributors
February welcomed 15+ first-time contributors, including @rodrigocoliveira (agno-client and multiple fixes), @ajshedivy (MCP HITL), @theagenticguy (AWS Bedrock features), @WilliamEspegren (Seltz tools), @dvy246 (Unsplash tools), @Tuyohai (Moonshot provider), @Byunk (Firestore fixes), @salimhamed (workflow serialization), @tanmaydarmorha (Azure auth), @Samyak2 (WebsiteReader fix), @mklrlnd424 (Tavily fix), @themavik (Knowledge and reasoning fixes), @ArivunidhiA (metrics fix), @guguoyi (docs), and @jss367 (link fixes). Every contribution strengthens the framework.
Community MVPs
- @aberk — Continued exceptional advocacy for production patterns: caching backend abstractions, error logging with stack traces, multi-worker setup support, and detailed feedback on v2.5 release. Always pushing for production-ready features with concrete technical proposals.
- @annedefined — Active community helper explaining framework comparisons (Agno vs LangChain/Langgraph/Pydantic AI), sharing production deployment experience, and helping newcomers with setup issues
- @kepler — Critical contributions including session state fixes and LearningMachine compatibility improvements, plus thoughtful questions about Learning Machines in Teams
- @rocasagrande — Raised critical production issue about media storage in DB (base64 overload), built comprehensive media_storage abstraction with S3 support, and actively contributed dependency injection fixes
- @Griflet — Continued advocacy for memory manager usage tracking and identified max tool call limit issues
- @Omar — Raised important production questions about image serialization and learning machine knowledge base behavior
- @LaGernouille — Detailed reporting on Gemini tool calling issues with thought signatures
- @josef — Proposed PR for cancelled runs being saved in DB
- @Ilya I. Lubenets — Asked important questions about dynamic prompt updates with Langfuse integration
- @chico — HITL for subagents in teams discussion
What's next
On the roadmap
- Studio enhancements: Visual builder improvements for complex workflows
- HITL for Workflows (When it's set on Agent/Team, which are used as a part of a step)
- Improved caching abstractions: Better support for multi-worker setups with pluggable cache backends
- Media storage optimization: Addressing base64 storage in DB for production deployments
- Enhanced observability: Better error logging with stack traces and debugging tools
- Workflow-level culture: Global context across workflow agents
- MCP Resources and Prompts: Expanded MCP protocol support
Get involved
We're always looking for contributors! Areas where you can help:
- Production patterns: Share your deployment experiences and best practices
- Documentation: Help new users get started faster
- Toolkits: Build integrations for your favorite services
- Bug reports: Production feedback makes Agno better (thank you to everyone filing detailed issues!)
- Frontend tooling: Help build better UI/UX for AI agents (check out agno-client!)
- Testing: Help test new releases and report issues
Check out open issues or join the Discord to connect with the team.
Join the movement
February 2026 proved that Agno is the production framework for building AI agents that ship. From Team Modes to HITL for Teams, from the Approval System to the Apache license change, every feature this month answered real production needs.
The community is thriving. 400 contributors strong. The projects are shipping to real users. The framework is maturing at an incredible pace.
Whether you're building autonomous AI marketplaces, voice-enabled RAG systems, or multi-paper research assistants, Agno gives you the production infrastructure to build AI systems that work reliably at scale.
Stats that matter:
- 400 contributors on GitHub 🎉
- 10 major releases in one month
- License changed to Apache 2.0
- HITL for Teams (finally!) shipped
- Hundreds of production deployments
- One incredible community
- ⭐ Star Agno on GitHub
- 💬 Join our Discord
- 📖 Read the docs
- 🚀 Share what you're building — tag us and we'll feature YOUR project in March!
Keep building, keep shipping, keep pushing the boundaries of what agents can do.
— Team Agno