Hi HN!<p>I’m Anton, a founding engineer at Athena Intelligence. Athena is an AI data platform+agent that supports workflows for enterprise teams.<p>Over the last two years we’ve deployed the state-of-the-art data LLM tooling to enterprise clients across several domains, zeroing in on co-piloting workflows and offloading more and more of the work to AI agents.<p>Today we’re proud to share a demo of the first (to our knowledge) autonomous data AI agent!<p>Athena engineering team has built several key pillars for end-to-end autonomous workflows:<p>* Data platform: Athena connects to all structured data sources and supports embeds of all popular unstructured document types. We're built on Apache Superset and Iceberg, balancing secure multi-database queries on client warehouses with lightning-fast data pipelines for complex workflows. Building an unstructured document platform in addition to that has enabled us to offer our users granular control over their knowledge retrieval pipelines. We support a broad scope of use cases, ranging from low-latency, high-volume simple retrievals with Groq Llama 3, to involved, self-reflective RAG pipelines powered by the most capable models (Clude 3 models, GPT-4T) and adaptive architectures.<p>* Execution platform: Athena has integrated SQL Editor and Jupyter Notebooks environments to work with data, Notion-like reports and Figma-like whiteboard to compile end knowledge work outputs, and Browse environment to explore and retrieve information from web. We are big believers in providing users with more UI options so they can experiment and find workflows that work for them. Most of our “apps” use Yjs backend, allowing live collaboration between people and agents.<p>* Headless and Human modes: every environment and its metadata are transparent to and can be manipulated by Athena (in addition to humans through UI). That means that any item of work can be collaborated on between AI and humans in a two-way fashion, in real time.<p>* Agentic tools and orchestration: each environment has a set of retrieval pipelines, tools, agents and agent orchestration items to support execution on specific tasks (example: find latest product catalog online) and end-to-end workflows (update our numbers and externalities, make a forecast for next months orders). We serve a wide range of models, both proprietary and open-source, through agentic architectures built with Langchain and Langgraph.<p>As the result, you have an agent that can:<p>* Intake real-world problems through existing workplace channels like email or Slack.<p>* Compile and iterate on a high-level plan to solve the problem.<p>* Execute the plan step-by-step, incorporating self-correcting and 'human-in-the-loop' behaviors.<p>* Access data sources, including freeform browsing and querying both structured and unstructured documents.<p>* Synthesize the final writeup for the multi-step analysis and disseminate it to platforms and colleagues downstream.<p>You can see the agent in action in the linked video.<p>The autonomous mode is in closed beta for now. We’re deploying it across select customers for the next month, battle-testing the architecture and taking notes. We're also excited about expanding and upgrading key parts of the architecture to unlock features such as:<p>* “Save checkpoints” agentic workflows to go back and forth between “execution paths” and specific nodes for a given workflow.<p>* "Time travel" capability with Apache Iceberg to enable versioning of data pipelines.<p>* Agent-to-agent interaction: have different agents collaborate on an analytical report to populate, review, and evaluate it before a human check is required.<p>* Asynchronous agents, autonomously executing background tasks like documenting newly connected datasets, identifying data drifts and compiling regular reports.<p>Excited to learn about other autonomous data agents out there and looking forward to your comments!