repository·main·Indexed 12 days ago
https://github.com/phidatahq/phidataA framework and runtime for building, running, and managing production-grade agent platforms. Agno provides an SDK for building agents, an AgentOS runtime for serving them, and a UI for management. It features a 'Capability Ladder' for incrementally adding tools, structured output, storage, memory, state, knowledge, learning, guardrails, human-in-the-loop, teams, and workflows to agentic software.
Agno is a framework and runtime designed for building, running, and managing agent platforms. It provides three main components:
Agno is designed to give developers control over their agent stack, including data, memory, and security (via JWT-based RBAC), while enabling a learning loop through simulations and usage data.
The 03_tools cookbook directory contains several examples demonstrating how to manage and control tool usage within Agno Team workflows. Key capabilities demonstrated include:
async_tools.py for asynchronous tool execution and async_toolkit_context.py to show how async toolkit tools appear in the team system message context.custom_tools.py.tool_call_limit.py.tool_choice.py.tool_hooks.py and member_tool_hooks.py.The Agent Skills cookbook provides examples for defining and utilizing agent skills and helper scripts. Available examples include:
basic_skills.py: Demonstrates fundamental skill implementation.sample_skills/code-review/scripts/check_style.py: Demonstrates how to implement code style checking skills.sample_skills/git-workflow/scripts/commit_message.py: Demonstrates how to implement git workflow skills, such as generating commit messages.Agno provides a comprehensive suite of features for production-grade agent platforms:
The cookbook/examples directory contains small, complete agent implementations that demonstrate different use cases:
Agno provides several types of hooks to intercept and modify the agent's lifecycle. The following patterns are available in the cookbook examples:
pre_hook_input.py): Run logic before the agent processes user input.post_hook_output.py): Run logic after the agent has generated a response.session_state_hooks.py): Hooks designed to read from or modify the agent's session state.stream_hook.py): Intercept the lifecycle of a streaming response.tool_hooks.py): Middleware-style hooks that wrap every individual tool call made by the agent.Agno provides several core primitives for managing user information, memory, and knowledge. These patterns allow agents to extract, update, and retrieve information to maintain context across sessions. The primary modes for these operations are:
Key primitives include:
Agno provides observability support for tracing and monitoring agents, teams, and workflows. You can integrate with various third-party observability platforms using OpenInference or direct integrations.
Supported root integrations include:
openinference or moving traces between projects.openinference or openlit.openinference.openinference.openinference.For complex architectures, there are specialized tracing examples for:
The cookbook provides specialized agents for different data types. Each workflow folder typically contains a basic.py (the smallest runnable example) and various task-specific variants.
_01_text_classification/: Assigning N labels (sentiment, intent, topic)._02_text_multilabel_classification/: Assigning subsets of N tags (hierarchical or flat)._03_text_extraction/: Converting text into typed Pydantic objects._04_text_span_labeling/: Marking character/token spans (NER, PII, evidence)._05_text_pairwise_preference/: Ranking A vs B (RLHF data shape)._06_image_classification/: Single or multi-label classification._07_image_extraction/: Extracting attributes, OCR, or captions into typed objects._08_image_bounding_boxes/: Region detection using (x, y, w, h) coordinates._09_image_extraction_to_vectordb/: Extraction, embedding, and storage for similarity search._10_audio_classification/: Clip-level labels (language, emotion, etc.)._11_audio_transcription/: Speech-to-text with diarization and timestamps._12_audio_extraction/: Extracting action items, attendees, or decisions from recordings._13_video_classification/: Clip-level labels._14_video_extraction/: Event detection, scene descriptions, and action timestamps._15_document_classification/: Classifying invoices, receipts, contracts, etc._16_document_extraction/: Extracting data from multipage PDFs into typed objects.The Agent Memory cookbook is organized into several functional areas:
01_agent_with_memory.py through 08_memory_tools.py demonstrate how agents use memory.memory_manager/ directory provides examples of using the MemoryManager API directly with PostgreSQL.optimize_memories/ directory contains examples of strategies for optimizing agent memories.Note: SurrealDB memory manager examples have moved to cookbook/integrations/surrealdb/.
Agent-as-Judge is an evaluation pattern where a model-based scoring system is used to assess the quality of an agent's output. This approach allows for automated, model-driven evaluation of complex tasks that are difficult to score with simple heuristics.
Available implementation patterns in the cookbook include:
agent_as_judge_basic.py).agent_as_judge_binary.py).agent_as_judge_batch.py).agent_as_judge_custom_evaluator.py).agent_as_judge_team.py and agent_as_judge_team_post_hook.py).agent_as_judge_with_guidelines.py).agent_as_judge_with_tools.py).eval_model detail key (agent_as_judge_eval_metrics.py).