May 7

Eduards Ruzga explores how the AI community debates local LLMs versus cloud APIs versus subscriptions – but almost entirely on vibes. He attempted to address this by building an open-source tool that calculates quality-adjusted tokens per dollar across all three options.
He shares surprising findings: running GLM-4.7 Flash locally on a $3,500 MacBook delivers 22×worse value than calling the same model via API. ChatGPT Pro provides 6.4× more tokens than Claude Max at the same price point.
The methodology – combining benchmark scores, hardware costs, subscription limits, and API pricing into one comparable metric – is open for scrutiny and contribution. Attendees leave with a live tool, a framework for evaluating AI cost-effectiveness, and concrete data to replace opinions with math.
Tool and source: https://wonderwhy-er.github.io/llm-value-comparison/
Full analysis: https://wonderwhy-er.medium.com/51fc0ad0dbd7

Complex agentic systems often chain dozens, or even hundreds, of sequential tool calls to complete a task. Near-instant resume performance is therefore a critical requirement for agentic systems running across multiple regions.
Traditional containers provide speed but not enough isolation. Full virtual machines (VMs) provide isolation but are too slow for ephemeral, per-task execution. When an agent needs to use a sandbox to execute code or access a file system, waiting seconds for a container or full VM to boot significantly degrades the user experience.
In this session, Vikram Vaswani explores the systems engineering behind Blaxel’s perpetual sandboxes – isolated compute environments that start and stop almost instantly while maintaining state between shutdowns. He explains how building an agent-native control plane using Kubernetes, Firecracker, and Virtual Kubelet enables sub-25ms resume times.
This architecture enables Blaxel sandboxes to start in under 25ms and remain on standby when not in use, eliminating cold starts without complex orchestration.
It also provides airtight isolation and secure compute environments, allowing agents to autonomously and safely execute arbitrary commands and code.

Key Takeaways

Understand how to build agent-native infrastructure on Kubernetes
Understand why cold starts severely degrade agentic user experience (UX) and throughput
Learn how sandboxes combine VM-level isolation with container-like responsiveness

Audience

AI application developers building or integrating high-frequency agents on Kubernetes
Architects and site reliability engineers (SREs) interested in patterns that balance performance, security, and operational simplicity at scale
Platform and DevOps engineers responsible for building Kubernetes platforms with strong isolation, low startup latency, and multi-region execution.

Generative AI agents can deliver impressive demos. But the transition from a compelling proof of concept to a reliable production system often fails—not because of scale, but because of trust.
In this session, Stanko Kuveljić presents a candid postmortem of a production-grade scheduling agent that exposed a structural vulnerability: a conflict between “politeness” and policy compliance. In attempting to be helpful, the agent overrode business constraints, revealing how probabilistic models optimize for plausibility rather than strict rule enforcement.
This talk explores what it means to operationalize generative AI responsibly. Rather than increasing autonomy, Stanko presents a practical approach he refers to as the concept of “Bounded Autonomy,” wrapping probabilistic models within deterministic control flows to enforce business invariants.
Attendees will learn:
Why large language model (LLM) agents drift toward conversational compliance over business policy, and how “helpfulness” becomes a production risk.
A practical architecture for embedding probabilistic reasoning inside deterministic control flows with non-bypassable invariant enforcement.
How to treat agent actions as testable behaviors instead of prompt outcomes, implementing deployment gates that reduce manual regression validation effort in production.
A reusable reference architecture and behavioral testing checklist that can be applied to agent-based systems.
This session is designed for leaders responsible for moving AI systems from experimentation to enterprise-grade reliability.”