AWS about teams in agentic world: why the organization became part of the AI architecture (Category AI4SDLC)

Reviewed by Stephen Brozovich of AWSA leader's guide to advanced team structures in an agentic world" It’s a good guide for leaders to change a company’s operating model when agents start working alongside people. Stephen's role is interesting. AWS Executive in Residence And he works with Amazon. 1999 years. During this time, he has gone from technical teams to working with people, culture and organizational transformation. And it's about transformation that he talks about in this report that he's targeting CIOs, CTOs, business leaders who need to understand how to change processes, teams and governance on top of it all.

Stephen proposes to look at these questions through four facets: economic, talent, structure, and governance.

1️⃣ Economics Stephen suggests we don't start with "build it all ourselves." For each workflow you need to choose the mode: use, compose or build

• Use - take the finished solution Compose – Collect on top of frontier models, data, and process
• Build - we teach or seriously customize our own. Build only makes sense where there is real business differentiation. Otherwise, a leader burns money before he understands his workflow.

2️⃣ Talent Here the report overlaps with the idea. expert generalist According to Martin Fowler and Thoughtworks, the value shifts from the person who writes the fastest with his hands in one stack to the person who understands the domain, sets the task to the agent, evaluates the result and stops the system in time. AI enhances not narrow specialization per se, but the ability to link domain, client, architecture, and result verification.

In this block, Stephen reflects on the Juns and shows the trap of a diamond-shaped organization. (few lords and junes, but many midles)Companies are cutting junes for fast AI ROI, paying more for senior talent and burning out future expertise. To perform new tasks, an inverted pyramid can work: 3-5 Senior engineers plus agents. But the whole organization must remain in the form of an hourglass, where there is a strong top, a thin middle and a living lower layer of learning. Otherwise 2034 Seniors will have nowhere to go. (||To be honest, this rationale is similar to well-meaning – it is aimed at the common good, but does not answer the question of how junes will help a particular company.||)

3️⃣ Structure Stephen has three types of structures.

• Model A - old IT ops: engineering built, then threw over the wall of operation. In the agent world, this breaks down: ticket culture kills context, operators don't see what the agent is doing, don't control the model and the data, and runbook doesn't cover non-deterministic behavior. - Model B - integrated subs: 3-5 Senior engineers create/own/responsible for running specific end-to-end workflows. The people who wrote the system prompts and tools of the agent themselves disassemble the incident. 3 nights. Less accountability, less loss of context. - Model C - Integrated Pod + Platform. On a small scale, the pod can survive on its own. Nana 10Different platform parts will begin to multiply: auth layers, observability stacks, guardrails and ways to burn the AI budget. Therefore, we need a common platform: runtime, memory, identity, observability, policy, cost controls. But the platform should not stifle pods’ autonomy, it should provide a safe road. Only Model C will work at scale, but Model B can start in a small organization.

4️⃣ Governance Here's Stephen tied up. Singapore Agentic AI Governance Framework and AWS AgentCore: Both agree on the same issues. Who is this agent? Who authorized it? What's he allowed? Is it working as expected? Can I do an audit?

I like the word governance as running infrastructure. Not a PDF with a policy, but a code that works on every request. Do not ask LLMs to “behave,” but to check accesses, identities, tools, and activities outside the LLM loop. That’s exactly the same logic I used in my posts. GitLab Act 2, agent-first IDP, Tailscale Aperture and Databricks production AISpeed without identity, policy, evals, observability and owners quickly turns into risk.

And here, the report overlaps well with my old organizational design topic. In Flo and Team Topologies I wrote that an organization is architecture, too: team boundaries, ownership, interfaces, latency communications. AWS adds the following layer: in the agentic world, the organization becomes part of the AI architecture. If the agent changes the execution rate, the leader must redesign not only the tools, but also the form of the team, the way people learn and the outline of responsibility.

For me, the main conclusion is this: in the coming years, companies that have an operating model that matures around AI faster will win, rather than just giving licenses to everyone who wants to.

#AI #AI4SDLC #Management #Leadership #Architecture #Agents