Agents now act on files, tools, browsers, and sandboxes.
But many systems still reduce the work to a chat transcript. That makes long-running tasks hard to resume, audit, branch, or compare.
Open-source agent infrastructure
Rath Team builds tools for agent systems that need real structure.
We work on OpenRath, a session-first runtime direction for agent workflows where tool calls, sandbox execution, branching, memory, and validation stay inspectable.
Start here
A clearer way to build agent workflows.
Make the session the unit of work.
A session should carry messages, tool evidence, runtime placement, validation results, memory reads, and workflow lineage through the whole agent process.
A runtime shape for inspectable multi-agent systems.
OpenRath explores session graphs, workflow composition, sandbox-backed execution, and examples that preserve evidence instead of just returning final answers.
Read the docs, inspect the code, then run examples.
The homepage points you to the practical entry points: documentation, repository, examples, notes, and official social channels.
Ecosystem
One place for Rath projects, writing, and code.
OpenRath
Session-first runtime for multi-agent workflows.
Repository
Docs
Quickstarts, API reference, and implementation guides.
Documentation
Blog
Research notes and engineering updates from the team.
Writing
Examples
Reference workflows for trading, engineering, and research agents.
Workflows
GitHub
Source code, issues, and public development activity.
Organization
PyPI
Installable Python package for the OpenRath runtime.
Package
Overview
Why Rath Team exists.
Agents need sessions, not copied chats.
We treat the session as the runtime object: messages, tool results, state changes, forks, and merges stay visible instead of being flattened into prompt history.
Workflows should be composed like software.
OpenRath experiments with Python-first modules, explicit handoffs, sandbox-backed execution, and reusable workflow pieces for research and engineering tasks.
Public notes make the work inspectable.
The docs, examples, blog, GitHub activity, and social channels give different levels of detail for builders, researchers, and people following the team.
Evidence flow
The useful artifact is the path of work.
For serious agent workflows, the final answer is not enough. Rath Team designs around the evidence created while the work happens: what was requested, which tools ran, where side effects landed, what changed, and which validation signal accepted it.
This gives builders a way to inspect, replay, branch, and compare agent runs instead of trusting a single response.
- Task Request enters a session
- Tools Calls attach arguments and results
- Sandbox Execution placement is recorded
- Branch Rejected paths stay visible
- Validate Tests or review close the loop
Use cases
Where session-first workflows matter.
Preserve task context, diffs, test runs, and validation results.
Track source material, intermediate notes, tool calls, and experiment state.
Keep policy constraints, action logs, risk checks, and repeated-run evidence visible.
Turn a run into an inspectable dossier, not just a generated summary.
Latest updates
What Rath Team is working on.
Team
A university-born team with industry execution experience.
Rath Team brings together builders from Tsinghua University and Sun Yat-sen University, with hands-on engineering experience from internships at leading technology companies. We combine research training with practical product and infrastructure execution.
The team has experience coordinating technical projects, managing cross-functional work, and turning early research ideas into concrete software artifacts. We also have committed initial operating capital, giving the project room to develop with focus.
Contact
Talk to Rath Team.
For research collaboration, open-source work, engineering questions, or partnership conversations, reach the team by email or through our public channels.