The number of applications a person uses to do their job is quickly collapsing to roughly one. If that’s right, there are only three places to stand, and many companies may be in the wrong one.
Really enjoyed this. Your Horse/Harness/Hay metaphor is cool but what really landed for me was a deeper argument about where durable power may sit in the agentic stack: context, workflow, learning loops, and ultimately ownership of the work.
I struggle daily with one thing I keep coming back to: The enterprise control side.
In many organizations, especially regulated ones like in healthcare and life sciences, systems of record are there not just to support work but to govern it, constrain it, preserve authority, and prove what actually happened within regulatory and organizational constraints. So the destination may be less about “maximum autonomy” and more “maximum useful agency inside a defined control envelope.”
So, that makes me wonder whether the system that increasingly understands and orchestrates the work will necessarily be the same system that is allowed to authorize, record, and govern it. Competence does not automatically confer authority, and escalation is different from approval, prohibition, and accountability.
Curious how you think about that tension. As agents take on more of the work, where do you see responsibility and institutional control ultimately living?
Where this lands hard for me: the harness you describe already has to hold more than one horse. Microsoft quietly shipped Claude models inside 365 Copilot next to GPT-5 — not as a hedge but an admission that no single model covers the whole job. So the ontology and learning loop in the Harness layer now include model selection itself, which arguably pulls part of the Hay (cost routing) up into the Harness. The founders I watch struggle at exactly that seam.
Very interesting article. But hard to follow some concepts. Like, "...does owning the next job make the last job better?" Is this a typo? That sounds like feedforward, not feedback. And, is the suggestion that a harness is a suite a good or bad thing? Hard to tell from the article; MSFT's lucrative success with a suite would suggest a good thing, but the article's tone hints at a bad thing. [shoulder shrug emoji].
Brett, this is a clear cycle map. I couldn’t resist taking the Brett test. You needn't feel complelled to grade it! ;)
I like that your piece names the layer we live in. Excipio is hay. Specifically, a private memory and control layer that audits and governs the traffic between every AI surface in an enterprise (apps, agents, tools, MCP servers) and the models they call. Governance and performance, at the primitives layer, entirely inside the customer's perimeter.
Your own-versus-rent question is the core fault line. Records to Salesforce, loop to the assistant. That split is the default trap unless the enterprise decides to own its context—and your third point proves why every serious organization eventually must own its context.
Nobody hands their operational intelligence to an external platform that also serves their competitors. Figma might disagree… But, the practical question is where that intelligence lives if it belongs to neither the horse nor the harness. Our answer: in a layer the customer runs on their own infrastructure. Nothing leaves the perimeter, nothing crosses tenants, and no intelligence leaks.
That is how knowledge compounds consistently instead of fragmenting across disparate surfaces. A correction made in one agent is immediately available to every other surface the next time it calls a model. The knowledge graph is built where actual traffic proves value, not mapped theoretically in advance.
On the four you told us to hold—Langfuse, LangSmith, Braintrust, Arize: they observe the call after the fact. We sit in the call. Controlling traffic inline provides the architectural guardrail, letting the same telemetry that measures a task grade it against an instrumented baseline. Your rule holds: sell the instrument, never the judgment. The customer’s definition of correct remains their own. We supply the proxy and the instrumentation, then get out of the way.
You asked what a hay company pools. Across the enterprise estate, we pool the telemetry between fragmented surfaces: which model routed which query best, at what latency and cost, and where reliability is degrading. No single internal team can see that across an entire enterprise, and no frontier lab will build it—because total transparency reduces redundant inference and makes the customer portable.
Two doors out of the harness layer. We walked through the one marked down into the plumbing, and we did it on purpose.
Brett, as a horse crazy dressage queen in my spare time, you had me at hello with this metaphor. Super well done.
I did remember that when I was using the analogy.
Really enjoyed this. Your Horse/Harness/Hay metaphor is cool but what really landed for me was a deeper argument about where durable power may sit in the agentic stack: context, workflow, learning loops, and ultimately ownership of the work.
I struggle daily with one thing I keep coming back to: The enterprise control side.
In many organizations, especially regulated ones like in healthcare and life sciences, systems of record are there not just to support work but to govern it, constrain it, preserve authority, and prove what actually happened within regulatory and organizational constraints. So the destination may be less about “maximum autonomy” and more “maximum useful agency inside a defined control envelope.”
So, that makes me wonder whether the system that increasingly understands and orchestrates the work will necessarily be the same system that is allowed to authorize, record, and govern it. Competence does not automatically confer authority, and escalation is different from approval, prohibition, and accountability.
Curious how you think about that tension. As agents take on more of the work, where do you see responsibility and institutional control ultimately living?
It's not just agents, all AI Surfaces are taking actions. I like your Heminway Idea that something could be "True at First Light, and a Lie By Noon."
Where this lands hard for me: the harness you describe already has to hold more than one horse. Microsoft quietly shipped Claude models inside 365 Copilot next to GPT-5 — not as a hedge but an admission that no single model covers the whole job. So the ontology and learning loop in the Harness layer now include model selection itself, which arguably pulls part of the Hay (cost routing) up into the Harness. The founders I watch struggle at exactly that seam.
Very interesting article. But hard to follow some concepts. Like, "...does owning the next job make the last job better?" Is this a typo? That sounds like feedforward, not feedback. And, is the suggestion that a harness is a suite a good or bad thing? Hard to tell from the article; MSFT's lucrative success with a suite would suggest a good thing, but the article's tone hints at a bad thing. [shoulder shrug emoji].
Brett, this is a clear cycle map. I couldn’t resist taking the Brett test. You needn't feel complelled to grade it! ;)
I like that your piece names the layer we live in. Excipio is hay. Specifically, a private memory and control layer that audits and governs the traffic between every AI surface in an enterprise (apps, agents, tools, MCP servers) and the models they call. Governance and performance, at the primitives layer, entirely inside the customer's perimeter.
Your own-versus-rent question is the core fault line. Records to Salesforce, loop to the assistant. That split is the default trap unless the enterprise decides to own its context—and your third point proves why every serious organization eventually must own its context.
Nobody hands their operational intelligence to an external platform that also serves their competitors. Figma might disagree… But, the practical question is where that intelligence lives if it belongs to neither the horse nor the harness. Our answer: in a layer the customer runs on their own infrastructure. Nothing leaves the perimeter, nothing crosses tenants, and no intelligence leaks.
That is how knowledge compounds consistently instead of fragmenting across disparate surfaces. A correction made in one agent is immediately available to every other surface the next time it calls a model. The knowledge graph is built where actual traffic proves value, not mapped theoretically in advance.
On the four you told us to hold—Langfuse, LangSmith, Braintrust, Arize: they observe the call after the fact. We sit in the call. Controlling traffic inline provides the architectural guardrail, letting the same telemetry that measures a task grade it against an instrumented baseline. Your rule holds: sell the instrument, never the judgment. The customer’s definition of correct remains their own. We supply the proxy and the instrumentation, then get out of the way.
You asked what a hay company pools. Across the enterprise estate, we pool the telemetry between fragmented surfaces: which model routed which query best, at what latency and cost, and where reliability is degrading. No single internal team can see that across an entire enterprise, and no frontier lab will build it—because total transparency reduces redundant inference and makes the customer portable.
Two doors out of the harness layer. We walked through the one marked down into the plumbing, and we did it on purpose.
Thanks for the positioning help!
Thats what I am here for!