Contact Homepage · pualpanwang@gmail.com
PaulClawX builds agentic tools for real-world workflows. Our work focuses on helping AI agents move beyond demos and operate reliably across browsers, desktop environments, interview training, and multi-agent collaboration.
Our direction is simple: move agents from “able to act” to “able to deliver.” Every run should be observable, recoverable, collaborative, and clear enough to diagnose when something fails.
| Area | What We Are Building Toward |
|---|---|
| Human + Agent | Agents that collaborate beside people instead of taking over the environment blindly. |
| Browser Agent | Web task automation with repeatable runs, traceable actions, and recoverable execution. |
| GUI / Computer Use | Agents that can understand and operate real desktop interfaces to reduce repetitive manual work. |
| Interview Agent | Structured feedback loops powered by rubrics, prompts, evaluations, and iteration. |
| Research Agent | Research workflows that search, synthesize, cite sources, and turn open-ended questions into usable briefs. |
| Multi-Agent Workflows | Multiple roles that split tasks, work in parallel, and consolidate results. |
| Tracing & Eval | Execution records that capture real behavior, failure modes, and improvement signals beyond polished demos. |
| Project | Mission | What It Focuses On | Status |
|---|---|---|---|
|
browser-agent Web workflow execution |
Never send a human to do a machine's job. | Browser automation, tool use, repeatable runs, action tracing, failure recovery, and infrastructure that turns page interaction into completed workflows. |
|
|
interview-agent Practice and feedback loops |
Turn practice into structured feedback loops. | Interview practice, rubrics, prompt iteration, evaluatable review flows, and agent collaboration patterns for training and long-term improvement. |
|
|
research-agent Traceable research outputs |
Turn scattered information into reliable research briefs. | Research task planning, web exploration, source collection, evidence-aware synthesis, citations, uncertainty notes, and decision-ready summaries. |
|
- Cute can still be rigorous: the experience can feel lively while the execution layer stays engineered and dependable.
- Autonomy should stay visible: actions, state, and errors should be easy to inspect.
- Failures are system inputs: the clearer the failure record, the more reliable the next run becomes.
- Human-agent collaboration is the default: people provide intent, judgment, and boundaries; agents handle repetition and workflow progress.
- GitHub: https://github.com/paulpanwang
- Email: pualpanwang@gmail.com