Great read, Eric. The seven-layer stack is a rigorous and much-needed piece of infrastructure thinking for the agent era.
One question that sits at the intersection of Layer 7 and the learning challenge: most of the framework describes technical and procedural controls that can be engineered, verified, and audited. Where do you see the organization's learning capability living in the stack? Governance and Lifecycle Management defines ownership and processes, but the harder problem may be whether the humans at that layer can actually learn from the mesh over time, reading patterns in agent behaviour, not just detecting explicit policy violations.
Related to Layer 6: you describe certification as version-aware and event-triggered, which fits well for discrete changes in code, models, or policies. Curious how you think about the slower, behavioural kind of drift, where operational telemetry gradually reveals that a certified agent's effective behaviour no longer quite matches its declared purpose, not through a single update, but through accumulated interactions, data shifts, or emergent strategies.
I suspect there is a rich essay or two hiding in that space, and would love to see you tackle it.
Thank you for the great feedback, and great questions!
Regarding where the organization’s learning capability live: I think HR (human resource) management practices will need to adapt in several ways; first we need to educate users and help them understand how to use agents in the most effective way; for agents, I think we need t o ensure that they have access to the correct business knowledge when needed (they need a “agentic knowledge fabric”), and we need to adopt HR practices (rubrics, measurement, optimization etc) for both people that use agents and for agents themselves; But still, it is a new field and I expect a lot of changes here.
Regarding “drift”; this is a big challenge today, but I come back to HR practices; to address “human” drift, we ensure people are continuously trained. Same for agents. Agents also need a much more comprehensive adaptation of HR practices (we call them AR, or Agent Resource management practices). This is the subject of a future article :-)
The AR framing is genuinely interesting, and I can see a rich essay coming from that direction.
I do wonder, though, whether HR and AR practices are necessary but not sufficient. In my experience, the harder challenge sits one level up: whether the organisation is actually structured to act on what the learning surfaces. You can have excellent measurement and training rubrics, but if the incentive structure rewards execution over inquiry, the learning loop runs as ritual rather than genuine adaptation. That is what I think of as the "lab vs factory" problem.
So perhaps the real question is not just how we manage agents and people better, but how we design organisations where the outputs of that learning actually change how the system is governed. Really looking forward to your future article on AR practices.
Agree that incentives belong in the rubric, and that is a good starting point.
One thing I keep coming back to is whether rubrics can reach the incentives that matter most, the structural and cross-cutting ones: who owns what risk, which metrics drive promotion and budgeting, how much organisational slack exists for inquiry versus throughput. Those tend to sit above any specific HR or AR practice and can quietly overpower even well-designed rubrics.
Curious whether you see AR practices evolving to address that structural layer, or whether the agentic mesh ultimately needs a broader organisational design pattern alongside it. This exchange is making me think there may be an essay or two on the organisational design side of the agentic mesh worth exploring further.
I want to ensure that Agentic Mesh remains focused on core technical stuff. But we are actively proposing “Agentic Process Automation”, “Agentic Resource Management” which may be the vehicle to address broader organizational issue. Good question, though, and I need to give it careful consideration.
That clarity helps a lot, Eric. It confirms the organisational design layer sits alongside the technical stack rather than inside it, which is exactly where I think some of the most interesting unresolved questions live. I will be following the APA and AR work closely, thanks for a really stimulating exchange.
Great read, Eric. The seven-layer stack is a rigorous and much-needed piece of infrastructure thinking for the agent era.
One question that sits at the intersection of Layer 7 and the learning challenge: most of the framework describes technical and procedural controls that can be engineered, verified, and audited. Where do you see the organization's learning capability living in the stack? Governance and Lifecycle Management defines ownership and processes, but the harder problem may be whether the humans at that layer can actually learn from the mesh over time, reading patterns in agent behaviour, not just detecting explicit policy violations.
Related to Layer 6: you describe certification as version-aware and event-triggered, which fits well for discrete changes in code, models, or policies. Curious how you think about the slower, behavioural kind of drift, where operational telemetry gradually reveals that a certified agent's effective behaviour no longer quite matches its declared purpose, not through a single update, but through accumulated interactions, data shifts, or emergent strategies.
I suspect there is a rich essay or two hiding in that space, and would love to see you tackle it.
Thank you for the great feedback, and great questions!
Regarding where the organization’s learning capability live: I think HR (human resource) management practices will need to adapt in several ways; first we need to educate users and help them understand how to use agents in the most effective way; for agents, I think we need t o ensure that they have access to the correct business knowledge when needed (they need a “agentic knowledge fabric”), and we need to adopt HR practices (rubrics, measurement, optimization etc) for both people that use agents and for agents themselves; But still, it is a new field and I expect a lot of changes here.
Regarding “drift”; this is a big challenge today, but I come back to HR practices; to address “human” drift, we ensure people are continuously trained. Same for agents. Agents also need a much more comprehensive adaptation of HR practices (we call them AR, or Agent Resource management practices). This is the subject of a future article :-)
The AR framing is genuinely interesting, and I can see a rich essay coming from that direction.
I do wonder, though, whether HR and AR practices are necessary but not sufficient. In my experience, the harder challenge sits one level up: whether the organisation is actually structured to act on what the learning surfaces. You can have excellent measurement and training rubrics, but if the incentive structure rewards execution over inquiry, the learning loop runs as ritual rather than genuine adaptation. That is what I think of as the "lab vs factory" problem.
So perhaps the real question is not just how we manage agents and people better, but how we design organisations where the outputs of that learning actually change how the system is governed. Really looking forward to your future article on AR practices.
Yes, incentives are a key factor. I see them established in the rubric, although it is early stage. Thoughts?
Agree that incentives belong in the rubric, and that is a good starting point.
One thing I keep coming back to is whether rubrics can reach the incentives that matter most, the structural and cross-cutting ones: who owns what risk, which metrics drive promotion and budgeting, how much organisational slack exists for inquiry versus throughput. Those tend to sit above any specific HR or AR practice and can quietly overpower even well-designed rubrics.
Curious whether you see AR practices evolving to address that structural layer, or whether the agentic mesh ultimately needs a broader organisational design pattern alongside it. This exchange is making me think there may be an essay or two on the organisational design side of the agentic mesh worth exploring further.
I want to ensure that Agentic Mesh remains focused on core technical stuff. But we are actively proposing “Agentic Process Automation”, “Agentic Resource Management” which may be the vehicle to address broader organizational issue. Good question, though, and I need to give it careful consideration.
That clarity helps a lot, Eric. It confirms the organisational design layer sits alongside the technical stack rather than inside it, which is exactly where I think some of the most interesting unresolved questions live. I will be following the APA and AR work closely, thanks for a really stimulating exchange.