🗺️ Roadmap: where TabAgent needs you

[dolev31]

TabAgent replaces an LLM agent's repeated closed-set decisions, such as which tools to shortlist or where to
route, with a cheap classifier trained on the agent's own successful runs. The paper is
arXiv:2602.16429. The library is pip install tabagent.

We've opened three tracks where outside help would make the biggest difference.

	Track	Start with
🔍	Find the bottleneck: which LLM call is worth replacing?	#1 OpenTelemetry traces · #2 LangSmith/Langfuse
🔁	Smoother hot swap: from head to production, safely	#4 LangGraph example · #5 shadow mode
🧩	A better feature card: what the classifier looks at	#7 candidate recency · #8 card explain

New here? Pick an issue labelled good first issue. Each one names the file to start in, what "done" looks
like, and the test to run. Comment on it with your plan to claim it (the Contribution proposal form helps), and we'll
answer within a few days.

The bar. Every pull request is checked automatically and reviewed against a published rubric:
CONTRIBUTING.md#the-bar. Run make pr-ready
before you push. AI tools and coding agents are welcome; a person stays accountable for every line.

Have an agent of your own? Run tabagent scan ./your_logs, then tell us in this thread what it found,
especially where it got the answer wrong. That's the most useful feedback we can get.

Ideas outside these tracks? Start a thread in Ideas.