For anyone else trying to sort out the difference, here's the example use case that came to my mind: Say you've developed an agentic workflow to review your inbox and update your Outlook calendar at the start of each day. It reads an email, identifies the date, time, and subject of a proposed meeting, and creates the appointment. Now say a client offers two possible meeting times. On the standard harness, the agent follows your pathing architecture. If you never accounted for the possibility of more than one option for a proposed appointment, you risk the agent quietly making the wrong choice. The way I avoid issues like this is to build in a hard stop-gate when unexpected hurdles crop up. The agent sees something it didn't expect, it asks the user to weigh in. The GitHub harness on the other hand might parse the rest of your calendar, weighing the long-running context of its interactions with you and selecting the time it determines is best. I'd want the standard harness here, because I want the final call. Others might want the more advanced harness because they prefer the efficiency over control and figure, "Hey, if both are open on my schedule then either works."
I reached that conclusion on my own after a lot of consideration as to what the article was actually trying to say. Much of the blog's explanation assumes a lexicon I don't have - 'blockers', 'publishing channel', and 'feature coverage' feel more standard for a reader who already understands much of the content at hand, rather than an amateur like myself. I've spent a couple of years and several hundred hours working with these models, plus a fair bit of coursework and actual development time with Cowork skills, yet I don't feel educated enough to glean any meaningful takeaways from the article. Maybe this information is simply intended for a more educated customer.
What I'd really like to know is what separates an organization that possesses a "readiness to govern usage-based billing" from one that doesn't. To me, readiness implies a willingness more than it does a certain degree of preparedness. In my job, for example, I've developed a billing skill package to allocate per-session credit costs internally. I'd think that meets the criteria, but I'm not totally sure because what constitutes readiness is not well-defined. In a world where many are spending thousands of dollars each month on AI-augmented work product, I'd really like to know how our protocols stack up against your metric so that we can make informed decisions about picking the right tool for the job.
Microsoft
