Saturday, August 1
Compass Stage
19 talks · 3 sessions
play_circle
Morning stream
play_circle
Afternoon stream
search
All
Keynote
Talk
Workshop
19 / 19 talks
AM
Session 1: AI Systems
00:00
The Limits of AI Coding Agents: Two Fundamental Gaps in Agentic Software Engineering
Ion Stoica
· Co-Founder, Databricks and Anyscale; Professor, UC Berkeley
AI coding agents haven't upended software and systems development because of two gaps that can only be narrowed, never certified closed — the requirement gap (user intent is broader than written requirements) and the model gap (the real world is broader than the development model) — and agents make both worse through context asymmetry, speed, and scale, which is exactly what reward hacking is.
Keynote
00:24
Photonics Is the Future of Computing
Nick Harris
· Founder/CEO, Lightmatter
AI's growth curve is outrunning the US power grid, and the real bottleneck is interconnect rather than compute; Lightmatter uses light to fuse thousands of GPUs into a single brain, approximating zero latency and infinite bandwidth to deliver 3× faster training and 11× decode interactivity.
Talk
00:33
Beyond Harnesses – Platform Solutions for Agent Reliability, Security, and Efficiency
Gosia Steinder
· IBM Fellow, IBM Research
Agent reliability, security, and efficiency are currently solved bespoke inside application-layer harnesses, but history (Unix/POSIX, containers/Kubernetes) says the industry consolidates onto a new operating system layer — and IBM Research is building that layer as an interception tier that pulls control functions out of agent business logic.
Talk
00:44
Is Kubernetes Good for Agents? Infrastructure Solutions for Agent-Shaped Problems
Tim Hockin
· Distinguished Engineer, Google
Agents are bursty, untrusted, single-tenant, and latency-sensitive — precisely the opposite of what Kubernetes was designed for. Google's open-source Agent Substrate doesn't rebuild Kubernetes; it layers an aggressively suspend/resume-based actor runtime on top, turning tens of seconds to tens of minutes of idle into sub-second wake-ups.
Talk
01:02
Why Your AI Agent Needs a Wallet: Agentic commerce on Arc with USDC and Nanopayments
Harshal Bhangale
· Staff Software Engineer, Circle
What stops agents today usually isn't model capability but payment — they hand control back to you the moment they hit a paywall or a signup flow. Give an agent its own budget-capped wallet plus x402 + USDC pay-per-call, and it can buy what it needs on its own.
Workshop
PM
Session 2: Frameworks & Dev Platforms
00:03
OpenSage: Next Generation of Agentic AI
Wenbo Guo
· Assistant Professor, UCSB
Every agent framework today makes a human pre-specify topology, toolset, and memory structure — which is exactly the "feature engineering" deep learning made obsolete a decade ago. OpenSage ships only a minimal scaffold and lets the agent spawn its own sub-agents, write its own tools, and design its own memory: what the speaker calls agent 2.0.
Talk
00:13
The Agentic Web and the Bazaar Era of AI
Ramesh Raskar
· Professor, MIT
AI is shifting from a factory era — a few large companies mass-producing what we consume — to a bazaar era of edge models running everywhere. Once billions of agents are on the open web, the decisive question isn't whose model is best but who builds the infrastructure, because whoever builds it will own it.
Talk
00:23
Toward Adaptive Agent Frameworks
Shiva Kasiviswanthan
· Principal Applied Scientist, Amazon Web Services
Today's agents plan and use tools well, but their execution policy is hard-coded. For mission-critical deployment, agents need to learn what to do next under an explicit budget (a constrained MDP), learn to explore several reasoning paths in parallel, and keep adapting to new tasks without forgetting old ones.
Talk
00:29
Information Retrieval in the Age of Agentic AI
Tanya Roosta
· Director of AI, AMD; Berkeley alumna
The unit of work in retrieval has moved from a lookup to an ongoing conversation with the world — yet we still grade the final answer with BLEU and ROUGE. Evaluation has to move to the trace: the correctness of every hop, whether each hop was adequate, and the cost in tokens and latency.
Talk
00:37
Are Small LLMs Ready for Coding Agents?
Denis Akhiyarov
· Senior Staff AI Scientist, ServiceNow
Shrink the harness to a single file, minimize context, take one tiny action at a time and verify it immediately — and a local small LLM really can do most of a realistic coding task. What it can't do is close the verification loop.
Talk
00:43
Agent Learning Requires Compressing Information into an Executable Reasoning Structure
Nilou Salehi
· Associate Professor, UC Berkeley
A model and a harness are doing the same thing — compressing large amounts of information into something executable. For enterprise processes that must run repeatedly with high consistency and explainability, the right executable form is a reasoning graph that an agent builds for itself.
Talk
00:49
Systems Foundations for Agentic AI
Johann Schleier-Smith
· Senior Staff Engineer, Temporal
What eats your time building agents isn't the core functionality but the cross-cutting concerns — retries, queuing, state management. Durable execution separates systems logic from business logic so you can write crash-proof agents in ordinary code. And more agency isn't always better: good abstractions let you slide up and down the agency spectrum without a rewrite.
Talk
00:59
Build AI Agents That Survive Failure
Nikolay Advolodkin
· Senior Staff Developer Advocate, Temporal
Put the agent's business flow in a workflow, wrap every non-deterministic thing that can fail — LLM calls, APIs, database queries — in an activity, and Temporal keeps the history, retries indefinitely, and resumes as though the failure never happened. Demonstrated live by pulling the plug on the weather API.
Workshop
PM
Session 3: Foundational Capabilities
02:00
Mindful Agents: Human-Inspired Memories for Long-Horizon Tasks
Doga Kerestecioglu
· Principal Applied Scientist, Microsoft Corporation
Agents have no storage constraint, and that is exactly why they never learned to distill; Microsoft's Fabric AI team borrows the human memory life cycle — ingest, consolidate, forget, retrieve, reconsolidate — to build memory for long-running agents over high-volume streaming data, and argues such systems need three distinct layers of evaluation: retention, retrieval, and task completion.
Talk
02:10
An Optimization Perspective on Recursive Self-Improvement
Ian Fischer
· Co-CEO, Poetiq
Real recursive self-improvement requires both the *recursive* and the *self* to hold; treating the LLM as one component of a broader system rather than the sole improvement target lets you build self-optimizing optimizers whose every RSI step costs one inference run instead of one training run.
Talk
02:21
Intelligence + Continual Learning = Expertise
Yu Su
· CEO, NeoCognition; Associate Professor, OSU
Intelligence (the capacity to solve problems) and expertise (accumulated, situated competence in a specific environment) are largely orthogonal axes; modern society is millions of micro-worlds each with its own local physics, which no static model can compress, so expertise has to be grown on the job through continual learning — and that axis, not raw intelligence, is the next dimension for scaling.
Talk
02:33
Reasoning as Control: Adaptive Test-Time Compute for Planning Agents
Furong Huang
· Associate Professor, University of Maryland
Self-improvement is usually framed as a model capability, but the model is only one layer; foundation models are increasingly runtime decision makers, and a self-improving agentic system can be improved at three levels — thinking (token-level decoding control), action (genuine self-reflection rather than imitating a stronger model), and workflow (retrieving a topology from a pre-computed bank at test time) — all without retraining the model.
Talk
02:45
On the Visual Capabilities of Multimodal Models
Raja Giryes
· Professor, Tel Aviv University
In today's multimodal models the reasoning happens almost entirely in the LLM while the vision side is a head bolted on, which explains their long lag on spatial cognition; but the two-year jump from near-chance to near-perfect on several tasks shows that asking the right question is itself what drives progress — and the mechanism turns out to be models internally *imagining* visual intermediate states.
Talk
02:57
The Open Agentic Stack: Building the Future of AI Systems with Open Source, Open Standards and Composability
Matt White, Ben Burtenshaw, Daniel Han Chen, Shang Yang, Romil Bhardwaj
· Matt White (Former Global CTO of AI, Linux Foundation; CTO, PyTorch Foundation); Ben Burtenshaw (Community Engineer, Hugging Face); Daniel-Han Chen (Co-Founder, Unsloth); Shang Yang (RadixArk); Romil Bhardwaj (Co-Founder and CPO, SkyPilot)
An agent is not a model, it's a stack — model, harness, tools, environment, guardrails, the RL systems that improve it, the evaluation infrastructure that measures it, and the compute it all runs on. Close and vertically control any one layer and the whole ecosystem becomes less portable, less competitive, and more fragile. Four talks work through four of those layers: environments (OpenEnv), models (Unsloth), reinforcement learning (Miles), and compute (SkyPilot).
Workshop
中文版