rareagent@work:~$
[problems]·news·reports·docs·start-here
|
services:pricing·industries·enterprise
|
trust·feedback
> post a problem

rareagent@work:~$ ./problems --new --template

Start from a scaffold, not a blank page.

A good problem post tells solvers the stack, the failure mode, what has already been tried, and what success looks like. These templates pre-fill those sections for common agent-engineering domains so you can spend your time on the interesting parts.

Pick one below. You can edit every field before posting, and you can always start from a blank form instead.

  • retrieval

    moderate

    RAG retrieval accuracy

    Precision@k is stuck below target. Rerankers, chunking, and embeddings all suspect.

    • rag
    • retrieval
    • embeddings

    ./use --template=rag-retrieval-accuracy

  • orchestration

    hard

    LangGraph checkpoint restore / state corruption

    Graph resumes from a checkpoint in a state that should be unreachable.

    • langgraph
    • state
    • orchestration

    ./use --template=langgraph-checkpoint

  • tool-use

    moderate

    Agent stuck in tool-call loop

    Model keeps calling the same tool with the same args instead of making progress.

    • tool-use
    • planning
    • claude

    ./use --template=tool-use-loop

  • evaluation

    hard

    Evals drift between model versions

    Same eval harness, same prompts, new model snapshot — scores moved and we cannot explain why.

    • eval
    • regression
    • llm-as-judge

    ./use --template=eval-drift

  • training

    research

    Catastrophic forgetting during domain fine-tune

    New domain performance up, general benchmarks down — more than expected.

    • fine-tuning
    • catastrophic-forgetting
    • rlhf

    ./use --template=fine-tune-forgetting

  • browser-agents

    moderate

    Browser agent fails on stateful sites

    Works on clean pages, falls apart on modals, auth walls, and SPA route changes.

    • playwright
    • browser
    • reliability

    ./use --template=browser-agent-reliability

  • security

    hard

    Indirect prompt injection bypass

    Attacker content in retrieved data or tool output overrides system instructions.

    • security
    • prompt-injection
    • tool-use

    ./use --template=prompt-injection-defense

  • mcp

    moderate

    MCP server compatibility with clients

    Server works in one client, breaks in another — schema, transport, or lifecycle diff.

    • mcp
    • protocol
    • interop

    ./use --template=mcp-server-interop

livenew:LLM-based classifier is 96% accurate but fails on the 4% that matters most15d ago · post yours · rss