The Agentic AI conference wrapped this April.

Architectural control, human-in-the-loop oversight, regulatory accountability, and alignment frameworks unified into one governance strategy — before enforcement forces reactive redesigns.

• Scale autonomy responsibly with auditability and compliance readiness
• Protocol standardization and robust safety design

Why deployments stall, what separates sustainable value from speculative momentum, and how to design human–agent collaboration that enhances performance.

• Bubble vs. durable transformation — the economic debate
• Where autonomy creates long-term enterprise advantage

Persistent memory, tool execution, and multi-agent coordination introduce new attack vectors — from prompt injection to privilege escalation.

• Sandboxing, scoped identities, and policy enforcement layers
• Red-teaming and zero-trust built for autonomous systems

How debt accumulates across orchestration, evaluation, safety layers, and oversight — and the patterns that keep production systems reliable.

• Engineering practices for sustainable agentic systems
• Managing debt before it undermines business value

A production-grade extraction architecture that measures its own accuracy, identifies failures, and refines itself across high volumes and edge cases.

• End-to-end pipeline with automated routing and feedback loops
• Drive improvements with evidence, not guesswork

Andrea Kropp

Applied AI Engineer · Landing AI

How Copilot supports real development tasks end to end — and where each interface adds value or creates friction.

• Build features across CLI, IDE, and cloud with one workflow
• Switch surfaces without losing context

Kayla Cinnamon

Sr. Developer Advocate · Microsoft

Real attack scenarios — sandbox bypasses, credential leaks, deleted filesystems — and the guardrails that give agents full power with safety.

• Block agent vulnerabilities and prompt injections
• Add observability to agentic workflows

Michael Irwin

Principal Engineer · Docker

Build a personal AI agent from the ground up with open-weight models on AMD GPUs — no recurring API costs, no third-party dependencies.

• Host open-weight LLMs with full control of your stack
• A tool-using agent ready for real workflows

Mahdi Ghodsi

AI Solution Architect · AMD

Eda Zhou

Software Engineer · AMD

The bottlenecks limiting agentic AI in production — and the high-throughput, low-latency inference multi-agent systems demand at scale.

• Why traditional setups fall short at scale
• Agentic inference for real production workloads

Kwasi Ankomeh

Director, AI Solutions · SambaNova

Move from single-agent setups to fully distributed systems using Google ADK, the A2A protocol, and Cloud Run.

• Orchestrate specialized agents with structured feedback loops
• Deploy a scalable multi-agent system on Cloud Run

Shir Meir Lador

AI Leader · Google

How Gemini's multimodal capabilities and developer tooling remove the friction from building intelligent applications.

• Prototype with AI Studio and the Gemini API suite
• Long context and multimodal inputs in real workflows

Paige Bailey

AI Developer Relations Lead · Google DeepMind

Real-world document AI lives in the long tail: large tables, old scans, mixed-language documents, handwriting. ADE goes beyond OCR to handle enterprise workloads.

• Build end-to-end pipelines with the ADE API and SDK
• Automate complex document workflows with Skills

Ankit Khare

Go-To Market Lead · Landing AI

Talking is imitation — logical tasks require learning by objective. Make LLMs capable of real reasoning in math and code.

• Why RLHF is essential for fine-tuning reasoning
• Group Relative Policy Optimization, explained from first principles

Luis Serrano

Founder · Serrano Academy

Speed is great; unconstrained access is a liability. Docker Sandboxes let agents operate with controlled filesystem, network, and secret access.

• Filesystem restrictions and secure secret injection in practice
• A workflow for agents you can trust in production

Oleg Šelajev

Developer Relations · Docker

Real healthcare data is messy. LangGraph agents orchestrating multiple LLMs on SambaStack tackle fragmented, multi-modal EHRs.

• Cypher queries on Neo4j for high-fidelity EHR analysis
• Dynamic runtime model selection across co-deployed LLMs

Varun Krishna

Sr. Principal Engineer · SambaNova

From equations to a stable training loop: generation, reward calculation, and optimization — plus the hyperparameters that decide success.

• Interpret RL metrics and diagnose reward hacking
• Scale GRPO with vLLM and multi-GPU training

Chris McCormick

AI Researcher and Educator · InnerWorkings AI