Today, many organizations are adopting agentic development, both to unlock its potential and to stay ahead of the curve. My current organization is no different. As part of this effort, a set of alpha teams are exploring its adoption, building early capabilities and sharing learnings to guide broader rollout.
Being part of one such alpha team, I have been observing an emerging pattern. Many teams are building similar capabilities (like PDLC orchestrators, agent workflows and supporting skills) but in slightly different ways, often tailored to their specific product contexts.
While this can feel like duplication at first, I believe it is actually driving rapid organizational learning. Sharing a few thoughts on why this phase exists and how we might navigate it more intentionally.
The Paradox: Standardization Needs Maturity
In mature engineering domains, we standardize because the patterns are well understood. With agentic development, we are still discovering the primitives:
- Evolving Problem Space: Moving from deterministic execution to probabilistic reasoning
- Forming Abstractions: Defining what an “agent” fundamentally is in our organizational context
- Emerging Operating Models: Especially how we handle “Human-in-the-loop” (HITL) handoffs
The Risk: In this context, early standardization doesn’t create a foundation instead it creates a ceiling. It constrains exploration before we know what is actually worth scaling.
The “Divergence” Phase: Learning at Scale
What we are seeing right now is a natural progression. It’s a phase characterized by:
- Parallel Experimentation: Teams building similar capabilities to solve immediate problems
- Local Optimizations: Moving faster by tailoring tools to specific team contexts
- The “Almost-Right” Stage: Multiple versions of the same idea, each slightly different
This is the “Broad Adoption” stage. It may look like duplication, but it is actually increasing our learning velocity. We are effectively running parallel A/B tests on architecture across the company.
The Real Danger: Fragmentation Without Direction
Divergence is healthy, but unmanaged fragmentation is not. The challenge arises when:
- Teams are unaware of parallel efforts
- Learnings are trapped in silos
- Solutions are too tightly coupled to be reused or migrated later
If we don’t have a path to converge, we aren’t innovating as effectively, we’re just drifting.
A Balanced Way Forward
To ensure this divergence leads to a stronger future state, I’m leaning into three guiding principles:
1. Visibility Over Restriction
We shouldn’t stop teams from building, but we should require them to share. Visibility through demos, shared registries or internal “RFCs” (Requests For Comments) allows the best ideas to gain natural gravity. It reduces “accidental” duplication while allowing “intentional” experimentation.
2. Standardize the Contract, Not the Tool
Instead of enforcing a single framework today, we should align on interfaces:
- Expected Outputs: What artifacts or checkpoints must an agent produce?
- Interaction Models: How does an agent request human intervention?
Aligning on the what allows teams to remain flexible on the how.
3. Modular “Build-for-Reuse” Thinking
Even in an alpha phase, we should avoid the “monolithic agent”. By keeping skills and orchestrators modular, we can ensure that when the time comes to converge, we can reuse the best components from different teams rather than rebuilding from scratch.
The “In-Flight” Reality: Our Journey
In our organization, we are currently in this “Go-Broad” phase. We are seeing this divergence play out in real time, with different teams exploring their own agentic implementations based on their context.
While it may look like multiple directions from the outside, from within it feels like a natural extension of the learning process where real-world constraints are shaping what works and what doesn’t.
My expectation is that convergence will happen in due course, potentially evolving into shared patterns similar to those described here. At the same time, this is still unfolding and we remain open to different paths as we continue to learn what truly scales.
Final Thought
One way I have started thinking about this transition is:
Enable divergence. Design for convergence. Execute with discipline.
We are still in an exploration phase and that is a healthy, if sometimes noisy place to be. The focus may not be to eliminate variation today, but to ensure that when convergence happens, it is grounded in real usage and shared learning.
If we continue to build, share and learn openly, the path toward a more unified approach should emerge more naturally.
.
Disclaimer: The views and opinions expressed in this article are my own and do not necessarily reflect the official policy or position of my current employer. This reflects a point-in-time perspective on a rapidly evolving field, intended to foster dialogue and shared learning within the engineering community.