Skip to content
View xxwtiancai's full-sized avatar
🎯
Focusing
🎯
Focusing
7 followers · 49 following
  • PolyU
  • SHENZHEN

Achievements

Achievement: Pull Sharkx2Achievement: YOLO

Achievements

Achievement: Pull Sharkx2Achievement: YOLO

Highlights

  • Pro

Block or report xxwtiancai

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
xxwtiancai/README.md

Xiongwei Xiao

Building agent applications and the systems behind them.

About  ·  Focus  ·  Open source  ·  Projects  ·  Contact

Conceptual Agent Harness architecture showing the agent loop, model, context and memory, policy checks, tools and sandbox, observations, traces, and evaluation

About

I build AI agent applications and contribute to the systems behind them: sandboxed execution, agent SDKs, tool use, and context management. My open-source work includes improvements to credential handling in OpenSandbox and proposed sandbox workspace protections in Google ADK. I am interested in the engineering details that make agents dependable in real workflows.

My longer-term interest is in agents that can build on previous experience. I am exploring agent memory and recursive self-improvement (RSI): how to preserve useful context, turn feedback into better behavior, and test whether an improvement holds up. Previously, I worked on video coding and visual computing.

Focus

  • Agent applications · Current work
    Building useful workflows around models, tools, and actions.
  • Sandbox & execution · Open-source work
    Credential handling, workspace boundaries, and controlled agent execution.
  • Agent memory · Exploring
    What should agents remember, retrieve, and carry forward across tasks?
  • Recursive self-improvement · Exploring
    How can feedback improve an agent's harness, and how do we verify the gains?

Open source

Selected contributions to the tools and infrastructure behind AI agents.

Project Contribution Stars
Strands Agents Harness SDK Preserve structured output during history offloading.
Merged · #5429		 ★ 8,998
OpenSandbox Harden C# and Kotlin SDK credential handling.
Merged · #2183 · #2372 · #2375		 ★ 16,278
MOMO CODE Keep graph execution slots busy as tasks finish.
Merged · #12		 ★ 1,107
Google ADK Contain relative paths within sandbox workspaces.
Open PR · #7700		 ★ 22,047
OpenHands SDK Expose MCP resources and prompts in the SDK.
Open PR · #6196		 ★ 1,322
Qwen Code Share compression cache for equivalent model selections.
Open PR · #15318		 ★ 28,821

Repository stars and PR status checked on October 30, 2026.

Browse my pull requests →

Projects

ActionAnything

An action execution layer for an agent harness: normalize and validate proposals, apply policy and required approval, execute permitted actions, and record redacted traces. Model calls and deployment isolation stay with the host application.

ActionAnything: host application to normalization and validation, policy and approval, executor, and result; policy outcomes and execution feed a redacted trace

Code & documentation →

nanoRSI

An auditable loop for improving a harness around a frozen model. Candidate changes are proposed, probed, contrasted, and composed from execution evidence; scoring happens only after freeze.

nanoRSI: frozen model and editable harness; propose, probe, contrast, and compose during evolution; freeze before judging; evidence recorded in a ledger

Code & documentation →

Contact

For conversations about agent applications, sandboxes, memory, or self-improvement:

Google Scholar

[email protected]  ·  [email protected]

Pinned Loading

  1. nanoRSI nanoRSI Public

    A minimal lab for improving programs, agents and model parameters. Real demos, frozen evaluations, inspectable evidence. Zero runtime dependencies.

    Python 3

  2. action-anything action-anything Public

    A model-agnostic runtime for safely executing, approving, recording, and replaying AI agent actions.

    Python 2

  3. QwenLM/qwen-code QwenLM/qwen-code Public

    An open-source AI coding agent that lives in your terminal.

    TypeScript 30.1k 3.4k

  4. google/adk-python google/adk-python Public

    An open-source, code-first Python toolkit for building, evaluating, and deploying sophisticated AI agents with flexibility and control.

    Python 22.2k 4.4k

  5. opensandbox-group/OpenSandbox opensandbox-group/OpenSandbox Public

    Secure, Fast, and Extensible Sandbox runtime for AI agents.

    Go 17.3k 1.6k

  6. strands-agents/harness-sdk strands-agents/harness-sdk Public

    Build an agent harness and control it end-to-end. Open-source SDK for production AI agents in Python & TypeScript - any model, any cloud.

    TypeScript 9.2k 1.5k