The thing that flipped my thinking was the token accounting. A bare "hello" is ~30 tokens on a plain API call and ~20,000 through OpenClaw, because the framework injects a ~7k system prompt, workspace files, a tool list, and schemas on every call. For a real autonomous agent that's a reasonable investment. For our data-discovery chatbot it was ~20k of context the model had to wade through to do something a 200-token prompt did better — so we dropped the framework. For our portal SRE agent we never reached for it at all; Cloudflare Workers AI did the bounded job without it.
The pattern we keep hitting: a lot of "agent" tasks are really a deterministic pipeline with one small LLM call where judgment is needed. The framework gives you flexibility you often don't need, at a token cost you always pay.
Not anti-OpenClaw — we still deploy it when the work is genuinely open-ended, and the playbook (https://autoclaw.sh) is there for that. Mostly curious whether others have landed in the same place, or found the opposite. Where has a full agent framework clearly earned its keep for you over a smaller setup?