Tao Xiang

Founder of JouleBeat · AI systems and agent infrastructure

Building reliable runtimes for long-horizon AI agents.

JouleBeat builds the harness layer that helps autonomous systems preserve state, verify progress, recover from failure, and operate within explicit constraints.

Previously built satellite-derived crop intelligence and algorithmic weather-risk systems. Now contributing production reliability fixes and ecosystem mechanisms across agent infrastructure.

01

JouleBeat: Public Systems

Public, inspectable systems demonstrating JouleBeat’s approach to state, verification, recovery, and constrained autonomous execution.

Autonomous software evolution

Evolution Kernel ↗

An explicit observe → plan → execute → evaluate → commit runtime for auditable, reversible codebase self-improvement.

  • Git-backed ledger, commit and rollback for every accepted change
  • Allowed-path enforcement, Firejail-based filesystem isolation and execution controls, and hard cost, token, iteration, and failure limits
  • Durable execution across restarts, with roles decoupled through inspectable files
  • Public v1.0: ~1,900 lines of Python runtime, 99 acceptance and unit tests, and four adapters
Autonomous research

DeepGraph ↗

An evidence-grounded research system designed to move from a research direction to checkable conclusions and reproducible experiments.

Implemented
  • Evidence maps, experiment plans, claim grounding, and provenance tracking
  • Experiment orchestration with multi-seed result handling
  • Deterministic evidence gates that can pass or block downstream output
Design principles
  • Negative-result preservation and traceable claims
  • Significance, leakage, and compute-matched baseline checks
  • Human review at consequential decision points
These are public technical systems demonstrating JouleBeat’s runtime principles. They are presented as inspectable engineering evidence, not as production-scale benchmarks.
02

From understanding complex worlds
to building systems that act reliably.

One accumulating capability stack: observe changing environments, predict uncertainty, and build systems that act reliably.

Observe

Commercial Earth Intelligence

Founded Hexun Technology and built independent crop-intelligence capabilities from commercial satellite data for agricultural monitoring and market decisions.

Documented operational contribution

During the MH370 search, Tao’s team conducted and publicly documented multisource satellite-image analysis, including hypothesis generation, historical-image cross-checking, and public volunteer coordination. Original report ↗

Act

Reliable Agent Runtimes

Leads JouleBeat’s work on explicit state, verification, rollback, failure recovery, and resource constraints for agents operating across long horizons.

Public systems

Evolution Kernel · DeepGraph

03

Open-Source Engineering

Selected merged changes, described through the production symptom, diagnosed failure mode, and engineering response.

Production symptomGemini overload errors did not trigger profile rotation or model fallback.
Root causeThe SDK sometimes rewrote the error without a leading “503,” bypassing the existing classifier.
ChangeAdded status-level and message-level classification, plus end-to-end regression tests.
Production symptomStrict OpenAI-compatible gateways rejected Fork & Chat requests with “Unknown parameter.”
Root causeInternal forkedFromIdentifier metadata leaked into the outbound model payload.
ChangeRemoved internal fork metadata at the API boundary before request serialization.
Production symptomRapid memory writes repeatedly queued the same parent directory for costly VLM semantic regeneration.
Root causeThe queue lacked a short-window identity check for semantic-parent work.
ChangeAdded a bounded 45-second deduplication window with regression tests.
04

Ecosystem Work

Creating repeatable interfaces between AI companies, hardware teams, developers, and regions.

Technical exchange

OpenClaw AI Ecosystem Sessions

Created and hosted a recurring technical exchange with sessions involving Kimi, Rokid, and MiniMax.

Mechanism & governance

A Reusable Community Process

Authored a public playbook covering applications, review, scheduling, sponsorship transparency, and conflicts of interest.

Cross-region reach

Bilingual Technical Channels

Built channels for technical exchange, interpretation, and developer participation across regions.

05

Background

Tao Xiang is a technology founder with a long-running focus on algorithms that must work in changing, uncertain environments—from observing the physical world, to predicting risk, to enabling autonomous systems to act reliably.

Trained in remote sensing at Wuhan University and in Information Engineering at The Chinese University of Hong Kong under Prof. Xiaoou Tang, he has built businesses and systems spanning commercial earth intelligence, commercial crop monitoring across major production regions worldwide, algorithmic weather-risk pricing, and agent infrastructure. Today he leads JouleBeat and contributes to open-source agent systems and cross-region AI ecosystem collaboration.

Hong Kong

The Chinese University of Hong Kong

Master’s degree in Information Engineering · Advised by Prof. Xiaoou Tang.

06

Selected Media

Selected third-party reporting on Tao and his previous ventures.

Current collaboration interests

Build reliable agent systems with JouleBeat.

Currently interested in conversations with:

  • Design partners with real long-horizon agent workloads
  • Model and inference partners working on reliability and recovery
  • Research collaborators in agent evaluation and harness architecture
  • Exceptional systems and research engineers

Particularly interested in workloads that run for hours or days, interact with changing environments, and require verifiable recovery from failure.