ShareXpace

One shared workspace. Less duplicated work. Fewer wasted tokens.

Comment

Inspiration

When several people or AI agents work on the same project, they often ask similar questions, repeat the same research, and waste time and tokens.

ShareXpace was inspired by one simple idea:

Before doing new work, check whether someone has already done it.

While multiple people or AI agents work on the same project/stuff, they probably ask similar questions, redo research, and end up spending both time and wasting tokens. ShareXpace grew out of a super intuitive idea: Before you kick off something new, check if teammates have already done it.

What It Does

ShareXpace is a platform of shared workspaces built for AI agents. Before an agent starts a task, it checks the workspace for similar questions or existing results. If there’s already a good enough answer, it just reuses it. If not, it calls the LLM and saves the result so teams can benefit from reusing it without spending extra token.

This helps teams:

  • Avoid duplicated queries/prompts
  • Share related&useful results
  • Reduce unnecessary LLM calls
  • Track estimated token savings

Estimated token savings can be represented as follows:

$$ T_{\text{saved}} = T_{\text{avoided input}} + T_{\text{avoided output}} $$

How We Built It

We built ShareXpace as an codex plugin so it can used by different codex user (i.e., AI agent)

Its main features include:

  • Check for similar tasks
  • Semantic search within embeddings vector DB to find related questions
  • Shared knowledge(i.e., storage) for questions and results
  • Source and timestamp tracking for reference
  • A dashboard showing estimated token savings and some relevant insights.

Codex helped us build and debug the plugin, APIs, and project architecture. GPT-5.6 helped generate the results for new question.

Challenges We Faced

The biggest challenge was deciding when two questions were similar enough to reuse safely. A high-confidence matched question should return the existing answer without LLM call, while a related question should use RAG, and a genuinely new question should trigger full generation via LLM call.

We also needed to perform the pre-flight check refrain from slowing down the agent workflow. Each request had to keep secure, workspace-isolated, and linked to the right AI agent and prompt.

In the last, we designed ShareXpace so RAG and full generation are handled by the user’s GPT model instead of requiring any shared but usage-paid LLM API key. This apparently reduced cost, but required the reliable handoff and result submission workflow.

What We Learned

I learned that reducing token usage is not just a caching-memory problem, it is also a teamwork and coordination problem.

Embeddings help to tell the questions with the same intent, while RAG lets existing shared knowledge useful for related team work. Most importantly, every question is not necessary to trigger a new LLM call. We also learned that a shared workspace for AI agents needs clear routing rules, reliable sources, and clear workspace isolation to make answers reuse both useful and safe.

AI agents become more useful when they can reuse shared knowledge or work history instead of starting from scratch every time.

What's Next

Next, I plan to improve ShareXpace with stronger access control, smarter or adaptively fine-tune similarity thresholds, automatic result validation/evalutation, and real token usage&saving tracking.

I also want to support more AI agent frameworks (e.g., OpenClaw, Hermes, etc), add task dependency graphs, and help teams understand which knowledge is reused most often. My long-term goal is to make checking shared knowledge as SOP before every LLM call.

Built With

Share this project:

Updates