Water

Water is an open source Python framework that the author describes as an agent harness rather than another agent framework. It does not supply the agents. It supplies the infrastructure around them: orchestration, retries and resilience, observability, approval gates, sandboxing and deployment tooling, so developers can concentrate on what their agents do. The design rests on three pieces. Tasks are the building blocks, each with a typed input schema, a typed output schema and an execute function. Flows coordinate tasks into workflows that can be sequential, parallel, branching, looping or a directed acyclic graph, and flows can be nested. The execution engine is an async runtime that runs tasks in order, handles errors, retries and checkpoints, and emits events. On top of that sit LLM tasks, streaming agents, multi agent orchestration, tool use, human in the loop approvals, sandboxed execution, a ReAct style agentic loop, layered memory and guardrails for content, schema, cost and topic. Resilience features include circuit breakers, rate limiting, caching, checkpointing and a dead letter queue. Integrations cover MCP and the agent to agent protocol, and there is a telemetry dashboard with cost tracking, execution replay, a command line interface and a playground server. It installs with pip and targets developers building multi agent systems who find existing frameworks verbose. It is free to use, and no paid plans are published. The listing points to the author's launch post, written by Manthan Gupta.