Talk
We Solved Building Agents. Now What?
The capability is real - Stripe ships 1,000+ agent PRs per week, Claude Code writes million-line codebases, Harvey serves 100K lawyers. But 95% of GenAI pilots show no measurable P&L impact. The industry's answer is 'the harness' - but what people call the harness is actually two things: a model-facing layer that fades with every model update, and an infrastructure layer that only grows. Most harness thinking is pointed at the wrong half. This talk breaks down the five primitives that make up agent infrastructure, who's building them, and why the composition layer is the real opportunity.
About
Hamza Tahir is a co-founder of ZenML, the open-source framework for robust AI workflows. He helps teams engineer reliable, portable pipelines—from classical ML to agentic chains—by decoupling workflow logic from the underlying infrastructure. Hamza has architected dozens of production lifecycles with stringent latency and compliance constraints, focusing on the critical "connective tissue" that makes or breaks AI in production. He is obsessed with solving the unglamorous workflow challenges: managing warm state, ensuring rollback safety, and maintaining granular lineage across complex, distributed executions.