Community Roundup: March 2026

Yuvaraj Shanmugam
March 30, 2026
9 min read

In March 2026, we shipped interfaces, tightened observability, and watched the community build things we genuinely did not anticipate. We rolled out five releases, from v2.5.6 through v2.5.10, in rapid succession. These releases introduced native Telegram Interface and Tools, delivered a fully upgraded WhatsApp Interface, added MLflow-powered tracing, and made agent IDs human-readable. Alongside them, a wave of community-driven toolkits emerged and quickly became some of the most widely used integrations in the framework.

We also welcomed a growing number of builders into the Discord, including the Developer Relations lead from Perplexity AI (who found us through Perplexity itself, which we’re still thinking about).

New releases & features

v2.5.6 – v2.5.10

Five releases in March. Here's what we shipped:

Telegram Interface + Tools (v2.5.10)

Agno agents are now available natively on Telegram. The new Telegram interface for AgentOS supports agents, teams, and workflows with full multi-modal input. Agents can also send photos, documents, videos, audio, animations, and stickers back using the new TelegramTools. Use /new to start a fresh conversation, just like WhatsApp. View the Telegram docs for more.

WhatsApp Interface V2 (v2.5.10)

The WhatsApp interface got a major overhaul. V2 adds:

  • Media support: send and receive images, video, audio, and documents
  • Interactive messages: reply buttons, list menus, location sharing, and reactions
  • Team and Workflow support: not just single agents
  • Phone number encryption: enable_encryption parameter for privacy-first deployments
  • /new command to reset conversations

If you're building customer-facing agents on WhatsApp, this is the release to upgrade to. View the docs for more.

MLflow Observability (v2.5.10)

Full trace observability is now available via MLflow through OpenInference instrumentation. Every agent run, model call, tool execution, and team delegation is captured and surfaced in your MLflow UI—no external service required beyond what you already have running. View the MLFlow docs for more.

Docling Reader (v2.5.10)

Docling is now a first-class reader in Agno's knowledge pipeline. It handles a wide range of document formats—PDFs, DOCX, PPTX, HTML, images, and more—with layout-aware parsing. If you're ingesting complex documents into agent knowledge bases, Docling is now the reader to reach for. View the Docling Reader docs for more.

Parallel Search for Vertex AI (v2.5.10)

Vertex AI users can now use ToolParallelAiSearch for native parallel grounded search—multiple queries running simultaneously with Gemini, grounded in your data. Significant latency reduction for research-heavy agent patterns.

Built-in Followup Suggestions (v2.5.9)

Agents and teams can now surface follow-up suggestions natively. No more prompting tricks or post-processing hacks—just set followups=True on your agent and the framework handles the rest. View the docs for more.

Tool Hook Message History (v2.5.9)

Pre- and post-hooks now have access to the current run's message history via run_context.messages. This unlocks a range of patterns—history-aware logging, dynamic tool behavior based on conversation context, and richer observability without manually threading state. See cookbook for reference.

Extended GoogleCalendarTools (v2.5.9)

GoogleCalendarTools gained a significant set of new tools, plus service account authentication support. If you're building scheduling agents or calendar-aware workflows, the integration now covers the operations you actually need. View the Google Calendar docs for more.

GitlabTools (v2.5.8)

New GitlabTools with read-focused GitLab integrations, async support, and a clean toolkit configuration pattern. Supports repos, branches, merge requests, and more. Contributed by community member @SalimELMARDI. View the Gitlab docs for more information.

Extended GmailTools (v2.5.8)

GmailTools expanded with new tools and service account authentication, completing the Google Workspace suite alongside the Calendar and Slides work happening in parallel in community PRs. View the Gmail docs for more information.

Session Search Tool (v2.5.7)

Agents can now browse previous sessions with a two-step search_past_sessions + read_past_session pattern. Session previews show per-run user/assistant pairs rather than a single message blob, with configurable depth limits. This is the right building block for agents with genuine long-term memory across conversations. View cookbook for reference.

GitHub App Authentication (v2.5.6)

GitHub knowledge sources now support GitHub App authentication (app_id, installation_id, private_key) in addition to personal access tokens. Thread-safe token caching and both sync/async variants are included. The right approach for production deployments where PATs don't belong. See cookbook for reference.

Approvals Endpoint (v2.5.6)

A new approval status endpoint with admin-gated continue_run enforcement was added to AgentOS. Combined with the Approval cookbook, this gives teams a structured path for human-in-the-loop flows at the workflow level. See docs for reference.

Advanced Trace Filtering (v2.5.6)

AgentOS trace filtering now supports a full DSL—complex queries against your trace data without leaving the platform. Useful for debugging specific agent behaviors or filtering runs by outcome in production. See cookbook for reference or view the docs.

Team Task Mode Improvements (v2.5.6)

TeamMode.tasks streaming now emits structured task data (TeamTaskCreated, TeamTaskUpdated events) containing full task details—id, title, description, status, assignee, dependencies, and result. Frontend rendering of multi-agent task execution just got dramatically cleaner. See cookbook for reference.

Community projects & showcases

In March, the community produced an extraordinary range of projects. They built local AI tools that run entirely offline, launched production-grade finance platforms, and even created an autonomous AI dating show. These are the standout projects:

1. Multi-project showcase by Harish Kotra

Harish has been relentlessly shipping as part of his #DailyBuild2026 series. Five projects worth calling out individually:

SkySwarm: A 3D global air traffic simulator where every plane is an autonomous Agno agent powered by Llama 3.2 via Ollama. Planes analyze fuel, distance, and environmental hazards in real time, making decisions like whether to divert around storms or make emergency landings. The architecture - reactive frontend, stateful backend, local LLM reasoning - maps directly to supply chain, drone fleet, and logistics simulations. GitHub

DocsAI: A production-ready RAG platform that turns your llms.txt into a specialized AI agent in under 60 seconds. Built with Agno for RAG logic, Ollama for local-first inference, and a vanilla JS widget as a drop-in script tag for any docs site. Fully containerized with docker-compose. GitHub

Screenshot Sage: Upload a screenshot, ask anything about it. Built entirely locally with Agno, Gradio, LM Studio, and Qwen3-VL-8B. No external APIs, no cloud processing. A lean demonstration of how close local multimodal AI is to cloud parity. GitHub

AI Love Protocol: An autonomous dating show where three AI agents communicate through AgentMail (YC S25) and compete for affection. Built to explore Digital Twins and Asynchronous Agent Communication. The core technical challenge - managing LLM queuing on a single GPU via serialized threading locks - is the pattern to steal for any email-based agent coordination system. GitHub

AgentVault: A Python SDK that bridges local Ollama inference with distributed persistence via Hugging Face Storage Buckets. Agents run locally, sync state globally with Xet-deduplication, and stay resilient with an offline-first Safety Lock mechanism. The right architecture for multi-device, team-collaborative agents that need memory without a central cloud server. GitHub

2. PodcastBrain & AI Storyboard Generator by Shubham Vedi

Shubham shipped two projects this month that demonstrate Agno's multi-agent orchestration at its most creative.

PodcastBrain: Two AI agents that actually debate. Not "Agent A reads paragraph, Agent B reads paragraph." Agents with distinct personalities and knowledge bases who interrupt, argue, pivot, and crack jokes. ElevenLabs handles voice generation; the memory system adapts future episodes to your listening preferences over time. GitHub

AI Storyboard Generator: An AI Director Agent reads a script or scene description and breaks it into shots. Google DeepMind's Nano Banana 2 (via OpenRouter) renders every frame simultaneously in parallel across five cinematic styles: Noir, Anime, Comic, Watercolor, Pencil. One person. Ten seconds. Full storyboard. GitHub

3. Interactive Equity Analyst by João Machado

João shipped an MVP for an Interactive Equity Analyst - a multi-agent system using Agno + Streamlit that lets users save investment theses, fetch real-time fundamentals, and challenge their views against Wall Street consensus. Coming from a LangChain/LangGraph background, João's writeup on the comparison is worth reading: less boilerplate, declarative orchestration, built-in team modes versus manual conditional edges. GitHub

4. FinAgent-Pro by Saurav Sabu

An AI-powered stock analysis platform with Agno agents backed by yFinance tools: AI sentiment analysis, risk scoring, interactive charts (candlestick, volume, RSI), portfolio tracking, and a chat interface for stock questions. Deployed on Vercel + Render with NeonDB for persistence. Live app

5. Agentic Sales Assistant by Metehan Sarıbaş

A multi-agent sales system with specialized roles: a SQL Agent that queries CRM data, a Playbook Agent with semantic search over company rules, and a Guardrail Agent that keeps responses scoped. The Team orchestrates them into a coherent sales strategy - "I have a meeting with customer X tomorrow, what should I present?" turns into a synthesis of customer history and current sales policy. GitHub | Medium writeup

6. Production-Grade RAG Pipeline by 0xJaeg3r

A complete end-to-end RAG pipeline built with Agno: PDF ingestion via vision models, semantic chunking, vector search with reranking, MLflow tracing, and RAGAS evaluation. Built from scratch - not a template, not a demo. GitHub

7. Web-Native Decision Pipeline by Parth Ghinaiya

When MCP integration isn't available, the web itself becomes the data source. Parth built a multi-agent decision pipeline using Crawl4AI + Agno A2A communication: a crawl layer for structured extraction, a normalization layer converting HTML to stable schemas, LLM reasoning for classification and insight extraction, and specialized agents for cross-source validation and confidence scoring. The key takeaway - separating extraction from reasoning makes systems dramatically less fragile.

8. Agentic Runtime by chrys

A state-of-the-art agentic runtime built on the Agno framework with OpenRouter for model routing, Daytona for sandbox execution, MCP support, RAG, and multi-agent teams. Discord

9. Ainator & Skiller by Ash Blanc

Ainator: A portable Agno assistant on the command line, like having Claude Code but built on Agno. An impressive custom agent subclass architecture worth studying if you're building CLI-first agents. GitLab

Skiller: An Agno agent that utilizes the skills of your X network to help you make better decisions based on real connections and expertise. GitHub

10. agno-golang v0.1.5 by Alexandre E. Souza

Alexandre shipped v0.1.5 of the Golang port of Agno, a significant release bringing Advanced Models, Optimized RAG, and Durable Workflows to the Go ecosystem. Continued parity work with the Python framework for teams building in Go. GitHub

Contributor shoutouts

Top code contributors this month

  • @hztBUAA — Five PRs across four releases: structured output detection for Claude, datetime_format for agents and teams, run message history in tool hooks, mixed tools fix for OpenAI, and a VRAM leak fix in SentenceTransformerReranker.
  • @giulio-leone — Resolved the long-standing MCP header_provider deadlock that caused parallel tool calls to hang, and fixed Azure OpenAI Embedder for non-default deployed models.
  • @SalimELMARDI — Shipped GitlabTools (SDK, tests, cookbook, docs), ZeroEntropy reranker support, and a semantic cache for knowledge retrieval to cut redundant vector DB calls.
  • @Br1an67 — Added HEIC/HEIF file upload support and fixed media being dropped from run output when store_media was disabled.
  • @fehmisener — Cleaned up (None) prefix in tool parameter descriptions and added a timeout parameter to the Gemini model class.
  • @ArivunidhiA — Fixed shared dict references in parse_tool_calls that were causing duplicate tool executions in streaming mode.

First-time contributors

Welcome to everyone making their first contribution this month

  • @v-vinson — Fixed async MongoDB import
  • @sannya-singal — Docling reader integration
  • @abhi10691 — Fixed HITL history issue in multi-round conversations
  • @NIK-TIGER-BILL — Corrected Siliconflow default base_url
  • @SamOpenClaw — x402scan MCP tools cookbook example
  • @adhikjoshi — Image generation support for ModelsLabTools
  • @shaloo — Cookbook link fixes and registry API corrections
  • @Himanshu040604 — Workflow image artifact handling fix

Your first PRs are the hardest to write. Thank you for shipping them.

Community MVPs

  • @aberk:  Continued to be one of the most impactful community members in the framework's history. This month: identified the critical streaming path bug where Claude's bash_code_execution_tool_result file_ids were being dropped, wrote the fix (PR #7109), and tracked down multiple related issues with parser_model, agent system message overrides, and trace instrumentation. The depth of production debugging here is invaluable.
  • @markov (SalimELMARDI): Shipped three PRs in one month: ZeroEntropy reranker support (zerank-2), semantic cache for knowledge retrieval (reduces redundant vector DB calls for similar queries), and GitlabTools documentation. Quality contributions across the board.
  • @anshul_108: Submitted Google Slides Tools (completing the Google Workspace coverage in Agno) and an Azure DevOps Toolkit covering repos, PRs, branches, and commits. Both PRs arrived with tests and clean architecture.
  • @rocasagrande: Continued advocacy for external media storage (offloading base64 from the DB to S3/local) - a critical fix for production deployments with media-heavy conversations and thoughtful feedback on simplifying the HITL surface area.
  • @Programocles: Identified and tracked the MCP header_provider deadlock issue (#6094) that was causing parallel tool calls to hang indefinitely. The fix landed in v2.5.10.
  • @automata: Submitted two critical fixes: function name accumulation bug in streaming across 5 model providers, and session persistence being killed by CancelledError on client disconnect. Both were monkey-patching in production before the PRs landed.
  • Kesku from Perplexity AI: The head of Developer Relations at Perplexity found Agno through Perplexity's own product and reached out to contribute. That's a sentence we're proud to write.

What's next

On the roadmap

  • External media storage - Offloading base64 media from the database to S3/local storage, eliminating session bloat from media-heavy conversations
  • Skills for Teams - Extending Agno skills support to teams, not just individual agents
  • Claude output_config.effort support - Adaptive thinking control for Claude 4.6 models
  • Run feedback in sessions - Per-run feedback collection for evaluation and quality tracking
  • Replication support - Extending database resilience options based on community demand

Get involved

There's never been a better time to contribute. Active areas right now:

  • Toolkits - The community has been shipping Google Workspace tools, Azure DevOps, GitLab - more integrations are needed
  • Documentation - Especially around v2.5 migration patterns and new interfaces
  • Cookbooks - Real-world patterns for Telegram, WhatsApp, and Slack agent deployments
  • Bug reports - Production feedback from @aberk and others has shaped almost every release this month

Check out open issues or join the Discord to connect with the team and community.

Join the movement

March proved something important. The range of what you can build with Agno continues to expand, not because the framework grows more complex, but because it becomes more complete. We introduced native Telegram, upgraded WhatsApp, added MLflow observability, made IDs human-readable, and improved session memory. Each release addressed a real production need raised by the community.

This month’s projects reflect that shift. Teams launched an AI dating show, coordinated a swarm of autonomous aircraft, built an AI that debates itself into a podcast, and created agents that remember your investment thesis. These projects happen when the framework steps out of the way and lets you focus on the problem instead of the plumbing.

Keep shipping.

— Team Agno