Just my two-cents as a very much boots-on-the-ground infrastructure engineer. We use LLMs heavily on my team -- to write code and review PRs. Our harnesses and guardrails are very very good so we get extremely high-quality output out of most models.

But getting the humans and agents to talk together is really fragmented right now. It's annoying that the agent with the context in my locally-installed Cursor can't really participate in a Slack conversation about a PR that 'it' principally authored (and which I've signed off on).

This is a real communication problem that we face every day, and eventually someone will solve it well and I will ask our leadership to give them a lot of money.