Full-stack engineering / AI agents

zhihao志豪

Building reliable systems for ideas that matter.

Full-stack engineer working on personal agents, persistent applications, and the infrastructure that keeps them running.

Explore my workGet in touch
01 / SELECTED WORK

Work

Development practice & architecture exploration.

01

Multi-user personal agents

Development practice
Agent runtimeContextTools

My backend work on a multi-user personal-agent service connects models, tools, and code generation. It includes generating HTML artifacts in isolated sandboxes, with continued attention to long-conversation context, attachment retrieval, and response performance.

Engineering questions
  • Define the boundaries between model calls, tool execution, and run state.
  • Use attachment references and on-demand retrieval to manage context size.
  • Trace time to first token and tool latency across the execution flow.
02

Resumable background tasks

System design
Async tasksCheckpointReliability

Complex tasks often outlast a single conversation turn. I focus on separating user goals, execution stages, and retries so progress stays visible, interrupted work can resume, and results persist.

Engineering questions
  • Model running, waiting, completed, and failed states explicitly.
  • Build checkpoints, idempotency, and retries into the task lifecycle.
  • Connect request handling and background execution through queues and event records.
03

From HTML to persistent apps

Architecture exploration
ArtifactsPersistenceSandbox

Building on HTML artifact generation, I am exploring a shared model for application identity, persistent data, and actions, so full pages, conversation cards, and background agent updates can work with the same application.

Engineering questions
  • Explore shared data, identity, and actions across full pages and dynamic cards.
  • Compare the execution boundaries of Cloudflare Workers and isolated sandboxes.
  • Design background data updates, usable-version switching, and failure recovery.
02 / ENGINEERING NOTES

Notes

Observations & open questions.

NOTE 01 · PERFORMANCE

Agent latency:
where does the time go?

Break a response into time to first token, context transfer, tool execution, and subsequent generation before deciding what to optimize. Tool schemas and historical attachments can affect perceived speed, too.

Measure first. Optimize second.
NOTE 02 · RELIABILITY

When a long task stops,
how does it resume?

Task state, checkpoints, and final results need distinct roles. Recovery means more than resending a request: duplicate execution, side effects, and concurrent progress all need clear boundaries.

Design with failure in mind.
NOTE 03 · PRODUCT SYSTEMS

Once an app is generated,
what keeps it running?

Beyond the page are data ownership, execution environments, and scheduling. I consider usable versions, recovery paths, per-app request budgets, isolation, and costs under real usage.

From generation to everyday use.
03 / EXPLORING NOW · 2026.10

Exploring

Research in progress · October 2026

EXPLORATION 01 · CODING AGENTS

Multiple coding agents,
working together with intent

Studying how a lead agent coordinates planning, task decomposition, parallel implementation, and code review. The focus is how coding harnesses handle workspace isolation, acceptance, and conflicts—not simply adding more agents.

Plan / Implement / Review
EXPLORATION 02 · CONTEXT & MEMORY

Long conversations,
continuous context?

Examining context, memory, and persistence boundaries between main agents and subagents. Exploring how message references and on-demand retrieval can restore relevant information after history is compressed.

Keep references. Retrieve when needed.
EXPLORATION 03 · SEARCH & TOOLS

Change the model,
keep the search layer

Exploring provider-independent web search and content-reading tools that can be reused across models. Weighing interfaces, source quality, and call costs before deciding whether a dedicated search agent is needed.

Independent tools. Interchangeable models.
04 / A LITTLE ABOUT ME

About

I see backend systems as part of the product experience. A smooth response, a task that is not lost, and a feature that stays useful all depend on engineering details.

When choosing technologies, I look at real execution traces, runtime state, and cost to understand why a problem occurs before choosing an implementation.

Agent engineering

LangGraph · DeepAgents · Multi-model integration · Tool calling · Context management

Backend & tasks

Python · PostgreSQL · Async queues · Outbox · State recovery

Apps & deployment

TypeScript · React / Next.js · Cloudflare Workers · E2B

Away from code, I enjoy cycling and strength training.

05 / LET’S CONNECT

Contact

Working on AI agents, backend systems, or developer tools?
I would love to hear about your projects, questions, and ideas.

Start a conversation by email