More about HKUST
A Survey of LLM Agent Serving Through the Lens of Context Engineering
PhD Qualifying Examination
Title: "A Survey of LLM Agent Serving Through the Lens of Context Engineering"
by
Mr. Yukun ZHOU
Abstract:
Large language model (LLM) agents do not solve tasks as isolated prompt
calls. They repeatedly read external state, invoke tools, update memories,
edit artifacts, branch into subproblems, and return to the model with a new
working context. The systems problem is therefore not long-prompt execution
alone, but repeated materialization of bounded model-visible context from
persistent, mutable, and partially shared state.
This survey studies context engineering as the source of LLM agent serving
workloads. We first organize agent-side techniques into three recurring
functions: retaining and compressing past state, selecting and grounding
external state, and coordinating reusable context across calls, branches, and
agents. We then translate these functions into serving-visible signals,
including context lifetime, volatility, provenance, materialization latency,
state validity, and branch structure. These signals induce three system
requirement families: context lifecycle management, critical-path
materialization of external state, and workflow execution for correlated
calls.
Finally, we use these requirements to interpret existing serving mechanisms.
First, KVCache management, long-context execution, and phase scheduling
reduce the cost of admitted tokens. Second, retrieval, tool-output
processing, and artifact views prepare the next context before inference.
Third, prefix reuse, prompt modules, cache handles, checkpointing, and
workflow runtimes connect repeated calls. This mapping shows that current
mechanisms cover important parts of agent serving, but rarely share a common
model of context as structured, versioned, reusable state with provenance. We
close with evaluation gaps and interface opportunities for exposing context
regions, validity dependencies, branch scope, and scheduling priority.
Date: Monday, 11 May 2026
Time: 10:00am - 12:00noon
Venue: Room 3494
Lift 25/26
Committee Members: Dr. Wei Wang (Supervisor)
Dr. Binhang Yuan (Co-supervisor)
Prof. Song Guo (Chairperson)
Dr. Chaojian Li