Software Architecture Superstream: Building Architectures for Agents (and Humans)

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

Introduction – Neal Ford (5 minutes)

Neal Ford welcomes you to the Software Architecture Superstream.

Building Evals for AI Adoption: From Principles to Practice – Mallika Rao (35 minutes)

Mallika Rao discusses the hidden risk of evaluation debt in production AI systems, drawing on her experience as an engineering leader at Twitter, Walmart, and Netflix. She explains why traditional metrics fail modern architectures, breaks down a five-layer evaluation stack spanning infrastructure and UX, and shares a diagnostic maturity model to help engineering leaders eliminate silent semantic failures.

Building Durable, Long-Running Autonomous Agents – Parminder Singh (35 minutes)

Most AI agents work in demos. Few survive in production. LLMs are stateless, infrastructure fails, context windows reset, and real-world objectives can span hours or days. Parminder Singh, cofounder of Redscope.ai, explores three pillars of durable agentic systems: execution, autonomy, and statefulness. Discover practical approaches for handling failures, maintaining memory and state, and escalating to humans without losing progress. Learn how to build resilient AI agents designed not just to respond, but to endure.

Break (5 minutes)

Architecting Agent Memory: Lessons from 30 Years of Chat Rooms – Jason Davenport (35 minutes)

Designers of chat systems have spent decades solving a problem that AI developers now face: how to organize, prioritize, retain, and retrieve useful information from an endless stream of conversation. Jason Davenport shows you how patterns from chat rooms, including threads, channels, event histories, summaries, search, permissions, and retention policies, can inform the design of memory systems for AI agents. You’ll see how to structure working context, short-term state, and long-term memory so agents can retrieve what matters without overwhelming their context windows or consuming unnecessary tokens. Expand your toolkit with practical architectural patterns for giving agents the context they need while controlling cost, complexity, and information loss.

Architecture as Code Case Study: Governing Agents in Consensus – Jonathan Johnson (35 minutes)

Architecture definition language (ADL) is a general pseudo-code that describes all aspects of software architecture: structure, constraints, governance, and a host of other guardrails. In this session, software architect Jonathan Johnson illustrates the use of ADL to specify code within specific architectural constraints and general concrete, empirical fitness functions to verify its capabilities. Using agents, he illustrates both halves of the loop: agents that generate a conforming Sysops Squad and write its fitness functions, then a consensus crew of local models that judges the deployed system against the same ASSERTs. Join Jonathan for this hands-on case study showing how to build governed agentic architectures beyond simple guardrails.

Break (5 minutes)

Architect the Boundaries, Not the Prompt – Sandhya Subramani (35 minutes)

Agentic systems need flexible reasoning, but they also need hard execution boundaries. Access controls, approval requirements, execution limits, and operational policies cannot just depend on whether the underlying model follows instructions. Senior developer advocate Sandhya Subramani takes you through an emerging architecture for placing deterministic boundaries around a nondeterministic agent. You’ll look at history-aware policy, where an action can be allowed or denied based on what the agent has already done; runtime intervention, where the application can redirect an agent before a bad decision becomes an action; and execution isolation, where access to files, commands, credentials, and networks is enforced outside the agent itself. Through a live demo using an open source agent harness, you’ll see how these controls work together around the agent loop.

Architectural Guardrails for AI Agents: Turning Decisions into Enforcement – Theo Valmis (35 minutes)

As AI coding agents take on more implementation work, software teams face a new architecture problem: Decisions that were clear to humans can easily become optional context for agents. Theo Valmis, founder of Mneme HQ, explores how to move architectural decisions from human readable documentation into a machine usable governance layer for agentic software development. He shows you the difference between retrieval and applicability, guidance and deterministic enforcement, and how decision authority, lifecycle, scope, and exceptions affect what an agent should be allowed to do. He also discusses the multi-agent problem, where architectural intent must remain consistent across coding tools and environments.

Closing Remarks – Neal Ford (5 minutes)

Neal Ford closes out today’s event.