[Video Podcast] AI Autonomy Is Redefining Architecture: Boundaries Now Matter Most - InfoQ

This episode of the InfoQ podcast explores why generative AI represents not just another automation layer, but a fundamental shift toward autonomy that is redefining software architecture. The conversation with Jesper Lowgren, Enterprise Architect Lead at DXC Technologies, reveals that traditional procedural workflows cannot contain AI's emergent behavior, and that the real architectural challenge is defining clear boundaries rather than controlling every step.

The Core Problem: Retrofitting AI Into Procedural Workflows

The discussion begins with a critical observation: 95% of AI proof-of-concepts fail, and the primary reason is trying to force generative AI into rigid, step-by-step procedural logic. Lowgren explains that this approach gives organizations all the costs of AI without any of the benefits, because autonomy and procedural logic are fundamentally incompatible.

"Once you turn on autonomy, you should expect unexpected behavior, and you cannot manage it with the same old procedural thinking," Lowgren states. The shift is from controlling the steps to controlling the boundary—defining what AI cannot do, what it is allowed to touch, what decisions it can make, and what goal it must achieve.

The Seven Boundary Dimensions

Lowgren introduces a framework of seven critical dimensions that define the boundary of an AI agent:

1. Scope - Understanding interaction points with non-agent systems
2. Goals - Defining intelligent goals that agents can pursue
3. Authority - Decision rights and what the agent can decide
4. Policy - Constraints and rules the agent must follow
5. Risk - Understanding and managing emergent behavior
6. Semantics - Shared meaning and ontology across agents
7. Evidence - Proving system behavior and maintaining records

These boundaries become increasingly critical as systems mature from simple assistants to multi-agent systems with autonomy.

Governance and Design Must Be Joined at the Hips

A key insight is that governance can no longer be an afterthought or separate layer. As innovation speeds up, you cannot keep governance outside and "catch up later" because these systems will drift. Governance and design must be built together from the start.

Lowgren uses the analogy of a merry-go-round spinning faster and faster—either you let go and fly off, or you move into the center where governance and design fuse together. This fusion is essential because AI systems will drift into "horrible things" if there's a mismatch between innovation and governance.

Practical Implementation: A New Design Process

Lowgren describes a radically different workshop approach where instead of humans designing processes on whiteboards, they define boundaries and let AI design the system within those constraints. In one example, a team defined boundaries around a call center process, then had AI generate an end-to-end design with 27 agents, which evolved to 33 after edge case testing by business experts.

This approach is "insanely fast" and represents a complete mindset shift from procedural thinking to boundary-driven design.

The Maturity Model and Evolving Guardrails

Lowgren outlines five maturity levels for AI systems:

• Level 1 (Ad Hoc): AI assistants with unmeasurable benefits
• Level 2 (Repeatable): Single-purpose agents with defined processes
• Level 2.5: Multi-agent systems without full autonomy
• Level 3: Multi-agent systems with autonomy requiring new operating models
• Levels 4-5: Speculative advanced autonomous systems

Guardrails evolve significantly between levels. For example, authority and decision rights become critical when autonomy is introduced, and the risk picture becomes much more complex.

The New Trade-offs: Drift, Debt, and Stability

Traditional technical debt doesn't apply to AI systems. Lowgren explains that when you introduce technical debt into AI boundaries, the system will drift and hallucinate. The new trade-off is about how much drift an organization can tolerate based on the criticality of the business problem being solved.

"How much technical debt can we afford?" becomes the central question. Less critical systems might tolerate more drift, while payment or trading systems require tight boundaries and minimal drift.

Responsibility Boundaries in the AI Era

Different architectural roles take on new responsibilities:

• Business Architects: Responsible for policy anatomy and structure
• Enterprise Architects: Hold the ecosystem view and ensure integration
• Data Architects: Critical for ensuring high-quality data and semantic consistency

Lowgren emphasizes that enterprise architects are not optional in this new world—they become essential for managing the entire ecosystem.

Advice for Developers: Think Systems, Not Tools

For developers currently working with traditional development lifecycles, Lowgren's advice is clear: "Get off it. It's a race to the bottom." Instead of trying to do traditional development better and faster, developers should invest time in learning about agentic systems and boundary-driven design.

He recommends starting with frameworks like CrewAI or Magento, but building systems using the boundary principles rather than procedural logic. The power of technology is changing things too fast for traditional approaches to remain viable.

The Cost Efficiency Dimension

An important practical consideration is cost. Frontier models like ChatGPT and Gemini are expensive to use at scale. Lowgren suggests that as agentic systems mature, there will be a shift toward using small language models for specific tasks within multi-agent systems, rather than having every agent hit expensive frontier models.

The Final Principle: Always Define the Boundary

The episode concludes with Lowgren's core message: "You remember the boundary, that's all." The boundary is the fundamental concept that enables safe, scalable AI systems. Without clear boundaries, organizations risk losing control of autonomous systems.

This represents a complete inversion of traditional architectural thinking—instead of defining what the system should do step by step, architects must define what it cannot do, what it can touch, and what goals it must achieve. The system then figures out how to achieve those goals within the defined boundaries.

The overall message is clear: architecture is more important than ever in the age of AI autonomy. Enterprise and business architects play a central role in shaping policy, boundaries, and system thinking so that AI systems scale safely and responsibly.

---

Key Takeaways:

• Retrofitting AI into procedural workflows is a fundamental mistake
• The shift is from controlling steps to controlling boundaries
• Governance and design must be built together from the start
• Guardrails evolve as systems mature from single agents to multi-agent systems
• Architects become more essential, especially enterprise and business architects
• Technical debt in AI becomes system drift—organizations must decide how much they can tolerate
• The boundary framework (scope, goals, authority, policy, risk, semantics, evidence) is critical for safe AI systems

The episode is available in both audio and video formats, providing practical insights for architects, developers, and business leaders navigating the transition to autonomous AI systems.