We are building the next-generation Agent Brain.
This isn't just another "wrapper around LLM APIs with a chat UI." It is a purpose-built Agent architecture designed for complex knowledge work. Our target scenarios focus on high-value domains like IP and R&D: these involve long-chain tasks, heterogeneous data sources, and strict evidence requirements, ultimately demanding verifiable outcomes rather than just conversational outputs.
You will be building a reusable Agent infrastructure and a reasoning/orchestration kernel. Your core focus will be solving three major challenges:
- How the Agent Runs — Execution Engine: The mechanics of the agent loop. This includes multi-step reasoning cycles, middleware pipelines, Planning & SubAgent orchestration, Checkpointing & state recovery, and execution controls (Permissions / Cost / Clarification).
- How the Model Thinks per Turn — Context Engineering: It’s not about stuffing the context window; it’s about organizing attention. You will solve core issues such as: input standardization, determining the exact capabilities and states exposed per turn, compressing long histories into an effective working memory, structured degradation under budget constraints, and normalizing tool outputs into traceable evidence.
- What the Agent Can Use — Capability Foundation: Sandbox environments (Docker / K8s), Memory Store, MCP Hub, Skills Engine, File System & Upload Pipelines, multi-tenant isolation, security, and observability.
If you have a long-term passion for transforming raw model capabilities into robust systems capable of reliably executing complex tasks, this role is a perfect match.