Liquid AI

Inspiration

Most AI agent frameworks require developers to manually define graphs, configure nodes, and handle failure logic themselves. We wanted to flip that. The idea was simple: what if you could describe a complex multi-agent workflow in plain English and have the system build, run, and fix it for you? DigitalOcean Gradient gave us the infrastructure to make that possible without managing our own model hosting.

What it does

Liquid AI is a text-based AI agent platform. You open a session, describe a task in natural language, and a Queen agent interprets your request and generates a multi-agent graph on the fly. Worker agents handle specific subtasks like web searching, documentation analysis, and summarization. Everything streams in real time so you can watch agents execute, make decisions, and call tools. If a node fails or produces poor output, the system detects the failure, evolves the graph, and redeploys improved agents automatically.

How we built it

We built the orchestration layer around a Queen agent that parses user intent and dynamically constructs directed agent graphs. Each worker agent is powered by DigitalOcean Gradient AI for inference. The event stream architecture lets users observe node execution, agent reasoning, and tool calls as they happen. We designed the failure detection and graph evolution system to capture errors at the node level and trigger automatic redeployment with improved configurations.

Challenges we ran into

Getting the Queen agent to reliably decompose ambiguous requests into well-structured agent graphs was the hardest part. Early versions would either over-fragment tasks into too many agents or under-specify the relationships between them. We also had to build robust error handling so that one failing node would not cascade and take down the entire workflow.

Accomplishments that we're proud of

The self-healing graph system works. When an agent fails, the platform does not just retry. It analyzes the failure, restructures the graph, and deploys a better version. We also kept the entire interface text-based, which means there is no drag-and-drop graph editor to learn. You just talk to it.

What we learned

Designing agentic systems that recover gracefully from failure is a fundamentally different problem than designing ones that succeed on the first try. We also learned that DigitalOcean Gradient makes it straightforward to scale inference without worrying about GPU provisioning, which let us focus entirely on the orchestration logic.

What's next for Liquid AI

We want to add persistent memory across sessions so agents can learn from previous workflows. We are also exploring letting users save, share, and remix agent graphs as templates. Longer term, we want to support multi-modal inputs so the platform can handle image and document analysis alongside text.

Built With

• css
• powershell
• python
• shell
• tsql
• typescript