Strong opinions, weakly held. There is one camp right now still arguing that all software companies go to zero. Separately, I like a lot of what DHH has to say about software you build for yourself; it’s a real niche with real uses, it’s super attractive. But I don’t think either is the main path for the enterprise. Take this with a pinch of salt as I sell into the space.
Collaboration works at scale
What I think tends to work at scale is collaboration. Across teams, across firms. Horizontal software with teams that have experience across the industry, that can actually be supported, and that carries some accumulated trust and reputation.
Building everything yourself also means maintaining everything yourself. Most enterprises aren’t structured to absorb that cost indefinitely, and they usually don’t want to. Winning will be about focus: scarce excellence in building the things that differentiate the enterprise, and buying the commodity. It still makes limited sense to build the commodity.
Keep the controls separate from the intelligence
On the technical side, harness and model separability feels directionally right to me. Keep the controls for intelligence separate from the supply of intelligence. We’ve had to solve versions of this problem with human colleagues for years; banks figured out a lot of the hard parts long ago. Agents should be treated more like colleagues than black boxes. Consumer tools can keep playing in the consumer space.
Headless operation is interesting too: something closer to a client/server model of interaction through a harness, where the parts that matter actually run on a different substrate with the controls you need.
The lock-in risk
What I’m watching is how quickly some people get locked into the context a particular tool provides. Instinct and Muse are early examples. That pattern could get ugly in the enterprise if it’s not thought through.
Enterprise is deliberately not winner-take-all. Satya has been clear on this for a while. You want the context and the substrate of action kept separate, not folded into one big dark agent that owns both.
Where vendors fit
A16z has a reasonable point that vendors to the enterprise remain one of the better places to go faster on AI when you can keep a known blast radius and real controls. We were at our AWS EBC recently and the consistent feedback from senior people was that ISVs are currently making the fastest progress on adopting new tech.
If you understand the blast radius and you’re careful about vendors (Pat Opet’s advice on this is useful), then collaboration can make sense. It stops making sense the moment it turns into large dark agents that aggregate your enterprise insights, your operational learnings, and your training tokens.
Curious where other people are landing.
Frequently asked questions
Should enterprises build or buy agentic AI software?
Focus on building what differentiates the enterprise and buy the commodity. Building everything yourself means maintaining everything yourself, a cost most enterprises are not structured to absorb indefinitely.
What is harness and model separability?
Keeping the controls for intelligence separate from the supply of intelligence, so the enterprise can govern agents like colleagues and change the model without losing its controls.
When does collaborating with AI vendors stop making sense?
When it turns into large dark agents that aggregate your enterprise insights, operational learnings and training tokens, rather than working within a known blast radius and real controls.