Beyond the Model: Agents, Verification & Control Planes
Talk 1: Dogwood: A new language for agent safety
We recently launched Dogwood, a new language designed for describing what agents can do, and when they can do it. Dogwood is open source, and you can use it today. In this talk I'll cover what Dogwood is and why we built it, cover some of the theoretical foundations (first-order temporal logic), and describe how Dogwood's scalable cloud backend compiles Dogwood to SQL and uses Aurora DSQL to build fast, scalable, stateful policies.
Talk 2: Evals Is the New Black (Box)
Evaluating coding agents is harder than it first appears. Picking a model and reasoning effort is only the beginning. Coding harnesses introduce additional variables through prompts, tools, connectors, memory, and cross-session state.
Reproducing a result is also difficult. Agents may discover prior solutions, inherit state from earlier runs, benefit from KV-cache reuse, or depend on tools whose outputs change over time. Harnesses and models evolve independently, so techniques that improve one combination may have little value for another.
This talk looks at agent evaluation as a systems problem. We’ll cover the practical challenges of isolation, reproducibility, measurement, and comparison, along with lessons from building an evaluation system for real coding workflows.
Thanks you Antithesis for making this event possible.
About Speakers
Marc Brooker is a VP and Distinguished Engineer at AWS. During his 16 years at AWS, Marc has worked on EC2, EBS, Lambda, and most recently lead the team that launched Aurora DSQL. He is currently focused on infrastructure for agentic AI, and the availability and security of our large-scale systems. Before AWS, Marc completed his PhD at the University of Cape Town.
Khawaja Shams is the CEO and co-founder of Momento. Previously, he led DynamoDB at AWS and was a scientist at NASA’s Jet Propulsion Laboratory. He received the NASA Early Career Medal for his work on the Mars rovers. Addicted to real-time data infrastructure and the Valkey ecosystem, he loves rigorous characterization of high-performance systems and AI models.
Daniela Miao is the CTO and co-founder of Momento. Previously, she led metrics at Lightstep, which was acquired by ServiceNow. She loves observability and instrumentation as tools for driving performance improvements. She is addicted to chasing 100 µs latencies for large payloads across the EC2 network.