We have been bitching and moaning of the lack of real outcome orientation in software for a decade now. I think now is the time to really make that happen. Without the clear notion of outcomes, anything that you do in agentic has a ceiling. That ceiling is your expression and the time it takes for you to define how stuff is ought to work.
Outcomes over outputs. That means: we define the outcome of the action, feature or solution rather define the artefact itself. The artefact (software) is the output. The outcome is what that artefact allows the user to do.
There are many kinds of outcomes. A solution often causes multiple outcomes to happen. I drink a cup of coffee (solution/action) and I get refreshed. Also: my tastebuds and brain has nice feeling. That's already two outcomes. Something changed in the real world. Not just bits flying in the internet. The user got their "job done" in the Jobs-to-be-done wording.
Action -> outcome. Solution -> outcome. Always ask for what is the outcome. The outcome allows you say more with less. And we will soon learn how important that is.
Babysitting the LLM and your agent
I've come to think of working with the LLM and the agent as value add and non-value add. There are times I can get the agent to make magic. Then there are times when I'm just watching closely and praying the agent does not do something stupid. And lo and behold, then the next mistake is written. I'm still much of the time babysitting.
But what am I really babysitting?
My watchfulness comes too often still from my catching the agent making the wrong solution. I can just see it. And I know the agent might find the mistake on a review pass. But it might not.
I'm all the time doing solution review on my work! I so see it now as I'm writing this.
I haven't actually spent the time describing the outcome I'm seeking. And then I know some of the outcomes I wish to make are such that my agent cannot observe. That makes me the passable / needs-work judge. I have to make the translation in my brain that this solution will not cut it for my users.
Organisations drowning in plausible-but-not-optimal
I so see it in front of me. Plausible solutions for any claimed problem are just one prompt away. Any agent can start charting out organisational life and those agents will hundreds of problems. Anybody can prompt for their preferred solution.
The conversation is stuck in my solution vs your solution. My idea is better than yours. But all these solutions are plausible. We can even ask agents to compare plusses and minuses.
The real optimal solution is always a tradeoff with user outcomes and cost. Comparing solutions without the reference of outcomes is just trading opinions. In many ways this has been the life in organisations for long. Now we just can drown our selves in options and opinion quicker.
Scarcity of right
LLMs and agents are making building software cheaper every day. The physical robots are doing the same for the physical side. The market will drown in plausible products. Stuff will be customised on the fly to your needs.
Yet there is still the "optimal for me". The product that fits my needs the best for the given price. Business strategy is all about finding your place and your differentiation. Your product must still be the best for the customer to get chosen.
The game is still about finding the way to be the most suitable for someone.
Tinkering with solution details is valuable. But that's not the thing still.
Outcomes -> options -> Hypothesis -> Knowledge
The right way is to actually try get to the outcomes the quickest. I know most of us know this by heart.
Name the outcome. Figure out some options. Make the hypothesis. If we make solution X then users will Y. The stuff we make may or may not have that effect. But then we know. Now we can make that cycle just fast. Cost of measurement is also nearing zero. There is no longer any excuse why not do it.
Let me finally come to the outcomes being the focal point.
Somebody says: we need to re-architect our system for coding agents. We discuss MCPs, data access and the likes. Claude can propose 10 different variants with ease. How we select: based on optimal outcomes for our users and our company. As measured by...?
An agent with a defined, measurable outcome doesn't need babysitting — it needs a scoreboard to ground itself. The agent must root itself into outcomes. Babysitting is what outcome-definition debt looks like at the session level.
There will be a time in the future, where I can just send my agent explore to get to outcomes. Then I will not be babysitting for wording on my user interface.
PS: 2 sentences in 1 paragraph written by ai. Spot those by the em-dash.