Hao Zhang
Research Scientist at ByteDance
Evolving agent harnesses and shared environments.
I develop AutoIA, a self-evolving agentic system that constructs and evolves agent harnesses (tools, skills, and execution strategies) and shared environments (data, memory, and context infrastructure). It learns from task feedback and past projects to improve its own construction strategies.
Research Directions
My research explores self-evolving agentic systems that continually improve agent harnesses and shared environments. The work combines a composable Context IR for heterogeneous data and derived context, agent-driven construction of reusable tools, skills, memory, and context infrastructure, and experience-driven evolution of the system’s own construction strategies.
The work connects agent memory and retrieval, test-time learning and serving, and semantic queries and vector search. It builds on my earlier work in hardware-efficient execution, graph systems and learning, and query optimization and distributed analytics.
My longer-term goal is multi-agent and human–agent collaboration at scale through trustworthy shared environments, so people and agents can build on one another’s knowledge and work over time.
Research perspective: Organizing the Shared Information World of Humans and AI Agents · The Next Generation of Context Management
Selected Systems
- AutoIA @ ByteDance — a self-evolving agentic system that builds and improves agent harnesses and shared environments. It uses a composable Context IR and task feedback to develop reusable context, memory, tools, and skills, while learning better ways to build them.
- SEMA — makes LLM-powered semantic operators first-class SQL primitives, enabling query optimization and adaptive execution. Its evaluation reports 2–10× speedups over three baselines on semantic-query workloads.
- GES @ Huawei — a production graph database combining a composable service architecture with factorized execution. It achieved audited LDBC SNB Interactive #1 results on the declarative (2024) and imperative (2025) tracks.
- TQEX @ Huawei — maps irregular SQL operators onto portable tensor runtimes across heterogeneous accelerators. Its evaluation reports 12.2× over DuckDB and 22.7× over HeavyDB on supported TPC-H SF100 queries.
Highlights
- 30+ publications in database systems and AI, including SIGMOD, VLDB, ICDE, COLM, EMNLP, ATC, TKDE, and The VLDB Journal.
- #1 on both tracks of LDBC SNB Interactive, the leading audited benchmark for graph database systems: Declarative, 2024 (3,000× over #2) and Imperative, 2025.
I received my Ph.D. from the Chinese University of Hong Kong, advised by Prof. Jeffrey Xu Yu and Prof. Hong Cheng, and my B.S. in Computer Science from Wuhan University.
Selected Honors & Awards
- ByteDance Soaring Star Talent Program
- Huawei Gold Award for Individual Excellence — Top 1%
- Huawei TopMinds Talent Program — First Class (highest tier)
Collaboration & Internship
I welcome conversations with researchers and industry leaders working on self-evolving agentic systems, context infrastructure, and next-generation AI platforms. I am especially interested in collaborations on self-improving agents, harness and shared-environment evolution, experience-driven agent-team construction, memory and retrieval, and graph/vector systems.
I am also recruiting research interns at ByteDance in Shenzhen; email zhanghaowuda12 [at] gmail [dot] com with [Intern] in the subject.