AI Superstream: AI Harnesses

Schedule

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

Introduction – Christina Stathopoulos (5 minutes)

Christina welcomes you to the AI Superstream.

Build Your Own Agent: Escaping the Harness – John Berryman (45 minutes)

“Agent harness” has become the default term for tools like Claude Code, Cursor, and Codex, but that term obscures the fact that the agent harness is the agent, and you can build one yourself. John Berryman, founder of Arcturus Labs, constructs a simple agent from first principles, showing just how approachable it really is. You’ll see how writing agent skills in plain English makes the agent’s behavior transparent, auditable, and correctable by you and by your customers. And he’ll discuss what’s on the horizon: agents that aren’t trapped in the terminal, that can be wired into any application, and that carry personal context wherever they go.

Procedural Memory Is All You Need: The Continuous-Learning Primitive Your Agent Harness Is Missing – Casius Lee (35 minutes)

Building an agent is easy. Keeping one alive in production isn’t. The layer that Casius Lee keeps coming back to is procedural memory, the part of an agent harness that lets it learn the job by doing the job rather than relearning it from scratch (and burning tokens) every time it runs. The Oracle AI developer advocate shares three failure modes he sees in production agent harnesses and discusses why procedural memory is the primitive that decides whether your harness survives in production.

Break (5 minutes)

Building Agent Harnesses to Match Frontier-Model Performance in RAG – Vineeth Kalluru (35 minutes)

Most enterprise RAG systems today plateau at retrieving documents and handing them to the largest model available, relying heavily on the model’s context window and system prompt to handle the rest. With recent advances in harness engineering, it’s possible to redesign RAG so that the harness manages search state, curates evidence, and verifies answers before leaving the model to do what it does best: generation. Vineeth Kalluru, generative AI solutions architect at NVIDIA, traces this evolution from single-shot retrieval to recent stateful, agentic search approaches, showing how the right harness lets a compact open model match frontier-model quality at a fraction of the cost. His talk includes a live hands-on tutorial based on a recently published research showcasing the concept.

The Agent Harness Interface: Code, Control, and Adaptive Systems – Nicole Koenigstein (35 minutes)

Once code is placed inside the agent loop, the harness has to do more than just wrap a model. It has to decide what to execute next, preserve useful state, expose the right tools, interpret feedback, and turn failures into corrective actions. That’s where agent systems become executable, verifiable, and stateful rather than just generative. Nicole Koenigstein, author and AI researcher, connects the idea of code as agent harness with adaptive agent systems. The focus is not on harnesses that rewrite themselves but on the interface through which agent behavior becomes observable and adjustable over time, from planning, memory, and tool use to feedback-driven control, model selection and effort, skill use, role assignment, and topology.

Break (5 minutes)

From Agentic Prototypes to Scaling in Production: Securing and Scaling the Harness – Sergio Gonzalez (35 minutes)

Moving a great agent prototype into a production environment means solving for many concurrent users, sensitive data, and unpredictable inputs and liable outputs. This transition requires a robust agent harness designed for security and resilience. In this session, Sergio Gonzalez, agentic engineering lead at Inception42, reviews the critical components of production-readiness, focusing on how the harness layer acts as the definitive defense against jailbreaking, manages real-time evals, and enforces agentic security boundaries. You’ll leave with a blueprint for a well-architected foundation capable of bringing agent solutions to prime time safely.

The Harness of Harnesses: Run Your Coding Agent Unattended, and Trust the Result – Jake Levirne (35 minutes)

Most agent harnesses help a model do the work. This one decides when the work is actually done. Jake Levirne, cofounder of SpecStory and former CPO of Docker, demonstrates deadreckon, an open source harness that wraps the coding agent you already use (Claude Code, Codex, Gemini) and lets it run a long task while you walk away. You write “done” in plain English and deadreckon turns that into checks that a separate watchdog process runs—a process the agent literally can't fool because the secret it signs off with lives outside the sandbox the agent works in. You get an auditable record of what changed, why it changed, and what it cost, not a chat transcript you have to take on faith. In this session, you’ll see the whole loop live: define done, walk away, and have more trust in the result.

Closing Remarks – Christina Stathopoulos (5 minutes)

Christina closes out today’s event.