The Reliability Stack: Engineering AI Agents that Actually Reach Production
Enterprise AI in 2026 has a production problem. The agentic systems that impressed in pilots are failing in the field — not because the models aren't capable, but because the infrastructure running them wasn't built for the complexity of real-world workflows. Multi-step tasks stall. State disappears mid-execution. A single API failure takes down an entire pipeline.
But engineering teams are finding their way through. A new class of infrastructure patterns is emerging—built around automatic failure recovery, persistent workflow state, and guardrails that keep agents on track when external systems behave unpredictably.
Featured Sessions:
The Guardrail Tax: Why Trust Is the Hardest Engineering Problem in Enterprise AI
At Mastercard, deploying AI isn't a velocity problem — it's a trust problem. As the Chief Data Officer of one of the world's most scrutinized payment networks, Andrew Reiskind sits at the intersection of agentic ambition and regulatory reality. He'll walk through why the line between co-pilot and autonomous agent has proven harder to cross than the industry expected — not because the models aren't capable, but because the infrastructure of trust required to run them at scale isn't keeping pace. Using the fraud use case as a concrete case study, Andrew breaks down where agents earn their autonomy, where humans still have to hold the line, and what it actually costs to build the guardrail layer that makes production deployment viable in a regulated environment.
No Shortcuts: How Merck Is Engineering AI Agents that Scale
At Merck, getting AI agents to production takes more than powerful models — it takes plumbing. As an executive responsible for nearly all of the infrastructure underpinning Merck's AI ambitions — multiple hyperscale clouds and, edge locations, an agent-to-agent and MCP gateway strategy, and new Google Cloud partnership — Sean Finnerty has a clear-eyed view of why most enterprise agentic deployments stall before they scale. The intelligence of the model doesn’t matter without the foundational infrastructure required to make agents reliable, compliant, and repeatable across an organization of 75,000 people. Using Merck's commercial AI initiatives as a live case study, Sean unpacks what production-ready agentic infrastructure actually looks like.
Why Stateful Orchestration is the True Battleground for Enterprise AI
Most enterprise agentic deployments don't fail because the model wasn't smart enough — they fail because the infrastructure underneath it wasn't built for the operational reality of long-running, multi-step workflows. As SVP of Engineering at Temporal, Preeti Somal has watched this pattern repeat across organizations: a promising pilot collapses in production the moment a system hiccup triggers context loss, a broken handoff, or a stalled pipeline with no path to recovery. The fix isn't a better model — it's separating the brain from the hands. Preeti breaks down what that architectural separation actually looks like, how leading organizations are building durable orchestration layers that keep stateful, self-healing workflows running under real operational pressure, and what it takes to move from fragile pilots to systems the business can actually depend on.
Featured Speakers:
At this event, we're bringing together the practitioners who have crossed that threshold to share what the playbook actually looks like: how they structured their agent workflows, where they drew the architectural lines, and what it took to go from a fragile pilot to a system their business depends on.