Every month I answer the same question by hand: what actually changed across LangChain, LlamaIndex, CrewAI and the OpenAI Agents SDK, and which of it breaks my stuff. 40 tabs. One hour. Stale in a week. I turned it into one agent config with Nimble. Full run:

The thing I had wrong about this category: a search API hands you back material to work with. You still read the twelve pages. a Web Search Agent takes the scoped research task and returns the answer, with the sources attached. Different job.

The whole agent is a config file. No logic in my script. input: the research task in plain language skill: "release analyst who reads changelogs, never infers a version" sources: prioritize the official repos, block the aggregators output_schema: the JSON shape of the answer

output_schema is the part that matters. Every change it reports has to carry a source_url. Required field. So an uncited claim isn't something I catch by eye. It's a schema violation. And because the shape is fixed, next month's run diffs against this one.

Two things it surfaced that I'd have missed: CrewAI 1.15.3 flipped tool-result caching to opt-in. It used to be on by default. Upgrade and it silently turns off. OpenAI Agents SDK v0.18.0 changed the default RealtimeAgent model to gpt-realtime-2.1. If you never pinned it, your agent changed models without telling you.

And the one that made me trust the output: LlamaIndex shipped nothing in the window. Latest release is 0.14.23 from 24 Jun, five days before the period even starts. It said that, instead of padding the answer. A negative result I can act on beats a paragraph of filler.

cron it on the 1st → open a PR against RADAR.md in your own repo alert only on type: "breaking" point sources at four vendors instead of four frameworks → competitive intel Same twenty lines of polling code.

The whole thing is four config fields and one research question. No code. Build a Web Search Agent for your own research workflow:

@nimble_search ty for share the info

@nimble_search thats sounds amazing

@nimble_search saved for tonight

@nimble_search ☝️ Always learning from the market.

@nimble_search Once you stop reading the twelve pages yourself, you also stop noticing when one of them is written at the model instead of at you.

@nimble_search 手工追这些更新，迟早被变更拖垮

@nimble_search Spanish startups need lean workflows, not forty tabs of stale updates. Automating framework tracking keeps founders focused on product and market fit.

@nimble_search Langchain has 40 tabs because it's a framework, not a product. the sdk is an api. stop building. start shipping.

@nimble_search Is 'block the aggregators' enforced at the fetch layer or just prompt text — because a page the agent reads can argue with a config file, and sometimes wins.