Omnigent is an open-source, Databricks-managed meta-harness - it sits one layer above your models and harnesses. Omnigent allows you to swap harnesses with ease so that teams can run, govern, supervise, and share agents in one place. We'll also share our observations and learnings from using Omnigent over the past few months.... Read more

Modern systems increasingly rely on isolation mechanisms—especially with the rise of AI agents and sandboxed execution—but there are many different approaches and they each have different strengths and tradeoffs. We'll cover: • Linux namespaces and cgroups • Node isolates and runtime isolation • Isolation methods commonly used in agent frameworks and harnesses • Containers and container runtimes • Virtual machines and KVM • Architecture-level isolation patterns • Practical tradeoffs and real-world use cases Sam has run production workloads at scale across Kubernetes, containers, VMs, lambdas, Node isolates, and sandboxing platforms like E2B, and will share lessons from deploying these systems in practice.... Read more

Tokens are the atomic units of data that large language and multimodal models read, write, and reason over to generate outputs. Token economics studies how AI consumption is metered, attributed, and connected back to business outcomes. Token consumption at large enterprises is climbing far faster than per-token prices are falling, and AI spend is increasingly becoming a CEO-level concern rather than an engineering line item. This talk examines token economics of custom-built agents, and the structural reasons behind their high costs. It surveys the emerging technologies and research to enhance token economics. Attendees will leave with a practical framework for locating and reducing their own token costs.... Read more

The July 2026 MCP spec RC drops the initialize handshake and the Mcp-Session-Id header. If you built session auth, audit trails, or agent kill switches assuming a persistent session existed, that assumption just broke — and most people running MCP-connected agents in production haven’t dealt with it yet. I rebuilt my own agent runtime against this change and will share the details of the shift in architecture with you all.... Read more

If you know how to run production applications, you already know more about LLM inference than you might think — until someone mentions KV caches, tensor parallelism, continuous batching, or prefill/decode and suddenly the stack looks completely different. This session maps familiar platform engineering concepts like scheduling, scaling, routing, observability, and resource management to the LLM inference world, while explaining the concepts that really are new. We’ll build and operate an LLM inference service along the way, leaving platform engineers with the mental model and vocabulary they need to start operating AI workloads confidently.... Read more

The talk is based on building secure execution infrastructure for autonomous agents at Incredibuild/islo and creating thousands of disposable environments in real workflows. It covers isolation, MCP/tool boundaries, short-lived credentials, snapshots/caching, traces, failure recovery, eval evidence, and why container-shaped assumptions become limiting once agents are long-running compute users.... Read more

Code review is treated as the final safety net of modern software engineering. That assumption is quietly breaking down as AI writes most of our code. In many teams today, LLMs generate refactors, tests, fixes, and migrations. The author reviews the code, runs tests, validates behavior, and understands the change end to end. Then they are still required to perform the sacred ritual of opening a pull request and waiting for approval. Days later, an async human review arrives from someone with less context and worse tooling. The result looks like rigor, but mostly adds delay and shallow feedback. This talk argues that traditional PR review has become theater in an AI-driven workflow. AI tools already catch style issues, missing error handling, security risks, and best-practice violations faster and more consistently than humans. Meanwhile, the failures that actually matter, performance, edge cases, operational impact, are rarely caught by late, asynchronous reviews anyway. Based on real production workflows, the session proposes a shift to AI-assisted self-review - engineers own their changes end-to-end, AI enforces consistency and standards, and human review moves to where it adds real value - design discussions, architecture decisions, and mentorship. If AI has fundamentally changed how we write code, it’s time to change how we decide it’s safe to ship.... Read more

Our tech lives are surrounded by AI. Everywhere we look: social media posts, videos, blog posts, and code. Companies expect employees to embrace AI, but at what cost? And what’s it doing to our mental health? While it’s satisfying to create with AI, it feels like we’re also losing a bit of our humanity and creativity along the way. In this talk Josh will look at how the world of technology has changed due to the AI boom, and how we can lean on our AI minions to produce meaningful, creative, non-slop work.... Read more

Enterprises are collecting large volumes of customer feedback, support interactions, and operational signals, but most teams still rely on manual analysis to identify patterns, risks, and improvement opportunities. In this session, I will share a practical enterprise AI workflow inspired by a real-world CSAT analytics use case, where AI was used to transform thousands of customer feedback responses into structured insights, leadership dashboards, and actionable signals. The talk will cover how to design an end-to-end AI workflow using trusted enterprise data, LLM-based classification and summarization, metadata capture, validation, human review, and self-service analytics. I will also discuss the lessons learned while moving from manual reporting to an operational AI pipeline, including data quality, prompt design, evaluation, governance, and how to build trust with business stakeholders. The goal of this session is to help data leaders, AI practitioners, and engineering teams understand how to move beyond GenAI demos and build AI workflows that deliver measurable business value in production environments.... Read more

Your agent passes your benchmarks. The customer still won't ship it. Inside an enterprise you don't own, an eval isn't a leaderboard score; it's the artifact that convinces a skeptical stakeholder to put an agent in front of their users and their data. This talk walks through building evaluations an enterprise customer will actually believe: sourcing test cases from the customer's own failure history instead of a generic benchmark, making each failure legible enough that a non-engineer can see what broke and why, and turning "looked good in the demo" into a documented sign-off. Drawn from Forward Deployed work at Notion and an eval background from Scale AI, this covers the tradeoffs and the cases where this still falls short.... Read more

An agent runs forty steps against a real repository. It reports success. Something breaks in production an hour later. Which step did it? For most agent stacks the honest answer is that nobody knows. The trace is a pile of model outputs describing what the agent believed it was doing, the run isn't reproducible, and re-running it produces a different plan anyway. Three properties fix this, and none of them require a better model. Reliability. The agent's world state comes from deterministic sensors, never from the model's own claims. When the model asserts "tests passed" and the sensor disagrees, that's a logged incident, not a state transition. Across an integration suite spanning six system archetypes, 611 of 611 injected false claims were refused, with zero false refusals on honest ones. Reproducibility. No RNG anywhere, ties broken lexicographically, plans replayed bit-exact from a hash-chained journal. Same input, same plan, every time: 120 of 120 generated systems deterministic, 114 of 114 crash-and-resume runs converging to the same state. Auditability. Every step writes an append-only record containing the full decision table that produced it. Tamper with any of it and the chain breaks (3,894 of 3,894 mutations detected). You can ask the system why it chose an action and get the actual reason, not a rationalization generated after the fact. Long horizons are where this pays for itself. At three steps you can eyeball the trace. At forty, determinism is the only thing standing between you and archaeology. I'll show LEDGER running against a real repository, break it on purpose, and replay the failure bit-exact. Then I'll be specific about the cost: what you give up, where the model still genuinely earns its place, and what this approach cannot do.... Read more

Most teams secure LLM applications the way they secure software: test before launch, add a guardrail, move on. That model breaks down here. Models get updated, prompts change, tools get added, and an agent that was safe last quarter has a different attack surface this quarter. This talk covers three gaps I keep running into in production AI systems, and what to do about each. First, why red teaming a model has to be continuous rather than a pre-launch gate, including what changed across iterations while testing NVIDIA's Nemotron content safety filter model with their AI engineering team. Second, guardrails. What content filters and safety classifiers actually catch, the failure modes teams assume are covered but aren't, and how to build an evaluation loop that tells you when a guardrail has quietly stopped working. Third, identity, the gap I think is most underestimated. Agents authenticate, hold credentials, call tools, and act on behalf of users, but the identity infrastructure we point at them was designed for humans making individual decisions. Sessions, MFA, least privilege, and audit trails don't map cleanly onto a non-deterministic system that spawns sub-agents and chains tool calls across systems. I'll walk through where that mismatch creates real exposure and what agent-native identity has to look like instead. Attendees will leave with a working model for continuously testing their own LLM and agent deployments, not a threat list.... Read more

Solving the black box explainability, traceability, determinism and real time learning problems. Demo of the next generation of language models.... Read more

Engineers default to the largest, best model. It gets the task done. But token consumption keeps climbing, agents retry, re-read, and loop, and the biggest model isn't always the right one. The usual tradeoff says pick two of speed, accuracy, and cost. This talk covers the holy grail: all three. I'll cover routing strategies between SLMs and frontier models, how to evaluate them, and real-world results from production systems.... Read more

While the industry chases agent reliability through smarter models, most agent failures trace to one design choice: letting the model decide what happens next. I built a software factory where the model only writes the code, while a deterministic controller drives Claude Code and Codex workers through the entire workflow, grounded by executable gates and engineering directives, with every transition, retry budget, and terminal owned by a versioned graph. I'll show the architecture, the failure classes it eliminates by construction, and what adversarial battle-testing taught me about where agent autonomy actually breaks.... Read more