Thursday, October 8, 2026
Scaling Agents From Benchmarks To Production
October 8 · 26 videos
Agents leave the lab. Safety is the priority. Memory systems evolve. Infrastructure scales. Security keeps pace.
“Your AI agent has no nervous system.”
Brains vs Hands: How to Run AI Agents Safely in Production — Viren Baraiya
Viren Baraiya · AI Engineer · 16 min
Watch on YouTube →Viren Baraiya explains why production grade AI agents require a separation between non deterministic reasoning and deterministic execution. He argues that treating agents as full scale applications rather than simple components is essential for safety.
- The Brain vs Hands framework separates the LLM reasoning from a durable, deterministic execution harness.
- Late bound Sagas allow agents to assemble idempotent tools into long running workflows that can span months.
- Applying the Single Responsibility Principle from microservices helps prevent model hallucinations in multi agent systems.
- Production agents must include human in the loop approval gates for critical operations like infrastructure management.
- Reliability is achieved by compiling LLM generated plans into structured workflows using tools like Netflix Conductor.
We Mapped 115 Microservices for Our Coding Agents — Kamalakannan Nandagopal, Postman
Kamalakannan Nandagopal · AI Engineer · 17 min
Watch on YouTube →Kamalakannan Nandagopal discusses how Postman uses an API Context Graph to provide coding agents with the necessary context for complex distributed systems. This approach bridges the gap between siloed repositories and large scale microservice architectures.
- Postman mapped over 115 microservices and thousands of endpoints to create a centralized API Context Graph.
- The system grounds agent data in static code analysis and production telemetry to ensure architectural accuracy.
- Autonomous agents generated a 21 page architectural review identifying technical debt that was previously hidden.
- Approximately 75 percent of pull requests touch APIs, making real time context updates critical for agent success.
- Moving from generic code search to graph based context layers prevents the exponential degradation of AI generated documentation.
Beyond the Benchmark: How Parallel Web Systems Builds Reliable Web Agents
Pray · Weights & Biases · 33 min
Watch on YouTube →Pray from Parallel Web Systems outlines a strategy for building reliable web agents by moving beyond static academic benchmarks. The focus is on creating an automated data flywheel that samples production queries to measure quality, cost, and latency.
- Parallel Web Systems uses Oracle Agents to generate reference answers for complex, high latency web research tasks.
- Component masking involves disabling specific tools during testing to ensure the LLM is not masking underlying bugs.
- The company prioritizes quality as the non negotiable metric, noting that fast or cheap agents are useless if they are inaccurate.
- Static benchmarks are viewed as marketing gimmicks that fail to represent the distribution drift of real world customer queries.
- An automated evaluation pipeline is necessary to maintain developer velocity as query diversity scales.
Turning Agent Memory Into Skills That Work — Will Lyon, Neo4j
William Lyon · AI Engineer · 18 min
Watch on YouTube →Will Lyon of Neo4j proposes using context graphs to transform amnesiac AI agents into systems that retain and reuse actionable skills. This method shifts agent design from fragile prompt engineering toward governed software engineering practices.
- The Context Graph Model integrates short term, long term, and reasoning memory using entity extraction and resolution.
- The Agent Procedural Execution protocol treats skills as typed execution graphs rather than static text.
- Skill distillation from memory graphs significantly increases task completion rates on industry benchmarks.
- The Neo4j Agent Memory Service monitors for skill drift to ensure grounding in regulated environments like healthcare.
- Persistent skills reduce token burn and latency by allowing multiple agents to learn from a single successful run.
How Cline Embraces an Open-Weights Model Ecosystem for AI Agents
Brian Armstrong · Weights & Biases · 40 min
Watch on YouTube →Brian Armstrong discusses the massive shift from proprietary models to open weights within the Cline coding agent ecosystem. He emphasizes that the agent harness is the key differentiator for reliability and cost efficiency in autonomous workflows.
- Open weight model consumption on the Cline platform grew 20x in four months due to a 90 percent cost reduction.
- Iterative Self Optimization uses frontier models to analyze failure patterns and automatically improve the agent harness.
- The platform achieved 10 point gains on Terminal Bench by moving toward interactive, multi turn evaluations.
- Enterprise adoption requires infrastructure stability and high prompt cache hit rates to make workflows economically viable.
- Agent neutrality allows users to plug any model into a high quality UI without being locked into a single vendor.
How Grammarly Scales Real-Time Inference with vLLM on CoreWeave
Kristoff · Weights & Biases · 23 min
Watch on YouTube →Kristoff discusses how Superhuman manages over 100 billion LLM calls per week by transitioning to unified fine-tuned models. This approach achieves sub-second latency and high availability through custom load balancing and dedicated infrastructure.
- Superhuman handles over 100 billion LLM calls weekly for 40 million daily active users with 99.99 percent availability.
- The system transitioned from complex multi-model legacy setups to specialized 1B to 4B parameter models for cost efficiency.
- A custom LLM Load Balancer manages traffic between internal Kubernetes clusters and CoreWeave to prevent single-cloud lock-in.
- Proactive background-triggered tasks generate 100,000 requests per second, placing extreme pressure on standard cloud ingress layers.
- Speculative decoding and quantization techniques help achieve costs significantly lower than using frontier models.
- Infrastructure as Code using Terraform ensures parity between internal and external inference environments.
Your AI Agent Has No Nervous System — Matt Gibiec, Dynatrace
Matt Gibiec · AI Engineer · 15 min
Watch on YouTube →Matt Gibiec explains why AI coding agents require real-time environmental context to avoid repetitive loops and production failures. He introduces the concept of a nervous system for agents to bridge the gap between local development and production reality.
- AI agents currently lack the sensory input needed to understand how services perform in actual production environments.
- Integrating observability data into the agent reasoning process prevents code changes that work locally but break dependencies.
- The Bluebox tool by Dynatrace injects production context directly into local development workflows and agent prompts.
- Human roles are shifting toward a supervisory capacity where they approve plans generated by context-aware agents.
- Organizations often prioritize deployment speed over operational metrics, leading to significant technical debt in AI implementations.
- Trust in production code serves as the primary metric for measuring the maturity of an AI-integrated development cycle.
Data Architecture for AI: From Training Pipelines to Agent Memory
Alon Horev · Weights & Biases · 38 min
Watch on YouTube →Alon Horev describes a fundamental redesign of data architecture to support the transition from batch training to real-time agentic systems. VAST Data proposes a disaggregated architecture to handle the massive memory and storage demands of modern AI clusters.
- The Disaggregated Shared-Everything architecture separates compute nodes from data nodes to enable linear scaling to exabyte levels.
- Offloading the KV cache from GPU memory to high-performance flash can increase token throughput by 10x for agentic workloads.
- Modern AI training jobs can consume up to 200,000 GPUs on a single fabric, requiring resilient data infrastructure.
- Agentic AI requires the ability to manage millions of ephemeral file systems and index trillions of vectors.
- Multi-tenancy and robust Quality of Service controls are essential for maximizing GPU utilization in enterprise environments.
- Unified platforms must allow data engine events to trigger real-time functions like immediate semantic search for video frames.
Why We Deleted Our MCP Server and Rebuilt It — Abhi Arya, Reducto
Abhi Arya · AI Engineer · 16 min
Watch on YouTube →Abhi Arya shares lessons from rebuilding Reducto's MCP server to prioritize architectural structure over simple prompting. He argues that agent-first software must focus on verifiability and clear responsibility for errors.
- Reducto processed 3 billion documents over three years and found that throwing frontier models at raw documents yields only 60 percent accuracy.
- The team deleted their original MCP server because providing a tool for every API endpoint led to incorrect agent workflows.
- The new architecture uses high-level tools that encapsulate structural patterns and expertise rather than raw API access.
- Implementing an ask don't guess policy and visible confidence scores helps manage agent uncertainty in production.
- A snapshot tool provides stateless agents with access to live pipeline states to improve decision-making.
- Organic adoption by non-technical teams like sales serves as the ultimate validation for internal agent tools.
NVIDIA Vera CPU: The Foundation for Agentic AI
Hannah · Weights & Biases · 23 min
Watch on YouTube →Hannah introduces the NVIDIA Vera CPU, designed to eliminate bottlenecks in agentic AI loops that require high single-threaded performance. This hardware shift addresses the latency issues inherent in traditional cloud-era chiplet architectures.
- The NVIDIA Vera CPU delivers 2x the performance of current x86 architectures by focusing on single-threaded speed for agentic loops.
- A monolithic die design using the Olympus core eliminates the high latency typically found in chiplet-based CPU architectures.
- The system features a 1.2 TB/s LPDDR5X memory subsystem and 3.4 TB/s of core-to-core bandwidth.
- Vera reduces memory power consumption to 40W compared to the hundreds of Watts required by traditional x86 systems.
- Perplexity reported 1.9x faster sandbox startup times when using Vera for bursty containerized workloads.
- The rise of agents shifts the CPU from a simple control plane role to a critical path performance component in AI factories.
Alastair Campbell & Michael Gove Disagree Agreeably on Brexit.
Michael Gove · The Rest Is Politics · 53 min
Watch on YouTube →Michael Gove and Alastair Campbell debate the long term consequences of Brexit and the ethics of modern political leadership. They explore how regulatory autonomy in AI and life sciences serves as a potential upside to the economic friction of leaving the EU.
- British businesses face 39.2 million customs declarations annually at an estimated cost of 7.5 billion pounds.
- Experts project a 6 to 8 percent hit to UK GDP as a direct result of Brexit economic friction.
- The Manchester Model illustrates how city level growth can be driven by strategic alignment with sovereign wealth funds.
- Manchester City faces 115 charges related to financial transparency and alleged sham contracts.
- Gove argues that the UK growth rate since 2016 has surpassed Germany despite the challenges of Brexit.
- Political leaders often embody a mix of genuine belief and cold calculation rather than being pure archetypes.
Expanding AI: Near and Far
Razmig Hovaghimian · Weights & Biases · 17 min
Watch on YouTube →This panel discusses how AI infrastructure is becoming a global equalizer for frontier markets and regional hubs. Experts explain why sovereign AI models and domestic compute reserves are necessary to prevent a digital divide.
- Firebird reserves 20 percent of its compute capacity for local domestic engineers and startups to build regional capacity.
- Public libraries serve as effective last mile delivery systems for AI literacy and community training.
- Data center investments in emerging markets can positively influence national credit ratings and macroeconomic indicators.
- The Nordic region is leveraging green energy for AI infrastructure while implementing heat reuse programs for local greenhouses.
- A recent AI hackathon in Armenia saw 500 applicants for only 20 available slots.
- AI is framed as a tool that can level the playing field for 50 billion dollar frontier markets.
Let Anyone at Your Company Ship Internal Apps with AI: Garrett Galow, WorkOS
Garrett Galow · AI Engineer · 17 min
Watch on YouTube →Garrett Galow explains how WorkOS enables non technical employees to build and ship internal applications using AI. The focus shifts from the ease of building to the difficulty of deployment and security in a corporate environment.
- WorkOS created a CLI called WOW and a serverless platform named Atlas to automate infrastructure and security.
- The company registry now contains 278 internal apps that generate over 3.3 million requests per week.
- Approximately 37 percent of the most active internal apps were built entirely by non engineers in a single commit.
- Cloud Days are used to empower non technical staff to lead the build process while engineers act as coaches.
- The most impactful tools often automate simple niche workflows like newsletter formatting or change log drafts.
- Security is maintained by making infrastructure secure by default through Okta SSO and Doppler secrets management.
How Jeetu Patel Runs Cisco Like the World’s Largest Startup
Jeetu Patel · Y Combinator · 43 min
Watch on YouTube →Cisco President Jeetu Patel discusses the company's massive pivot toward AI infrastructure and the implementation of founder mode at scale. He explains why AI agents will drive a 450 percent increase in bandwidth consumption compared to human users.
- Cisco secured 9.3 billion dollars in AI related infrastructure orders over the last two years.
- AI agents consume 450 percent more network bandwidth than human users which creates a supply constrained environment.
- The Rule of Thirds for leadership involves balancing scale operators, market thinkers, and acquired founders.
- Patel argues that unlearning past experiences is as critical as learning new skills during major market shifts.
- Improving by just 1.27 percent every day results in a 100x improvement over the course of a single year.
- Currently less than 2 percent of the global population are considered power users of AI agents.
Building the Cloud for AI Agents | AWS CEO Matt Garman
Matt Garman · a16z · 56 min
Watch on YouTube →AWS CEO Matt Garman outlines the transition to a cloud architecture optimized for autonomous AI agents rather than human workflows. He details a massive 220 billion dollar capital investment plan to support the next generation of AI demand.
- Amazon plans to deploy 220 billion dollars in capital investment for 2026 to expand AI infrastructure.
- AWS aims to purchase 2 million NVIDIA GPUs over the next two years while reserving capacity for startups.
- Custom silicon like Trainium and Graviton can reduce costs by 20 percent while improving performance for high scale inference.
- Security must move to machine speed to handle AI powered vulnerability detection and prioritization.
- Startups are viewed as the leading indicator for enterprise technology trends five years before they reach mainstream industries.
- Infrastructure constraints are shifting from GPU availability to power, cooling, and high bandwidth memory.
Giving AI Agents Memory That Learns: Jake Broekhuizen, LangChain
Jake Broekhuizen · AI Engineer · 16 min
Watch on YouTube →Jake Broekhuizen explains how to move beyond simple observability to create AI agents that actually learn from experience. He introduces a memory taxonomy to transform stagnant trace data into durable context for future runs.
- Procedural memory accounts for approximately 75% of visible behavioral gains in AI agents.
- The Read-Write Cycle framework uses a four-stage loop of Read, Run, Filter, and Write to enable long-term learning.
- Developers should distinguish between working memory for temporary execution and long-term memory for durable data stores.
- Manual review of agent tone and compliance is unscalable in regulated sectors, necessitating automated signal extraction.
- Human-in-the-loop review is recommended for changes to procedural memory to protect core brand behavior and prevent regressions.
Built Together: Inside the Dell + CoreWeave Engineering Partnership Powering Next-Gen AI
Sarat Krishnan · Weights & Biases · 27 min
Watch on YouTube →This video explores the engineering partnership between Dell and CoreWeave as they co-develop massive rack-scale AI systems. It highlights the shift toward software-defined infrastructure to manage the extreme complexity of next-generation hardware.
- The pod is becoming the new unit of the computer, integrating compute, networking, and storage into a single managed entity.
- A Shift Left strategy moves hardware diagnostics and firmware stabilization from the data center back to the factory level.
- The NVL72 rack features an external NVLink switch with 1,296 individual connections to handle high-density compute.
- Large-scale Coolant Distribution Units now require up to 3MW of power capacity to support liquid-cooled AI systems.
- Early access to hardware allows for platform stabilization and firmware tuning that provides a significant runtime advantage.
Essentials: Manage Stress & Build Resilience | Dr. Elissa Epel
Dr. Elissa Epel · Andrew Huberman · 30 min
Watch on YouTube →Dr. Elissa Epel discusses the biological impact of stress and how to build resilience through cognitive reframing. She explains the difference between threat and challenge responses and their long-term effects on cellular aging.
- Chronic stress and rumination significantly dampen mitochondrial enzyme activity and accelerate cellular aging through telomere shortening.
- A challenge response increases cardiac output and oxygenation, leading to faster physiological recovery than a threat response.
- Radical acceptance involves stopping the energy drain caused by wishing unchangeable situations were different.
- Daily restoration is a critical predictor of long-term health outcomes, including depression and mortality risk a decade later.
- Engaging in generative and purposeful challenges has been shown to lead to hippocampal growth even in elderly populations.
AI Tokenomics Explained
Shri Koparkar · Weights & Biases · 34 min
Watch on YouTube →Shri Koparkar explains the evolving economics of AI inference as workflows shift from simple chat to high-volume agentic systems. The session details how to optimize token utility and infrastructure to scale AI profitably.
- Agentic AI workflows typically use 15x more tokens than standard chat interactions due to multi-turn tool calls.
- Agentic workflows offer a 98% cache hit rate compared to only 60% for traditional chat interfaces.
- Right-sizing models to specific tasks is the most effective lever for reducing token costs and improving performance.
- Only 8% to 16% of current AI use cases actually require premium frontier models for successful execution.
- Organizations should shift their primary KPIs from tokens per second to tokens per watt and cost per task.
Psychology of Addiction (And How to Beat It Without Willpower)
Rob Dial · The Mindset Mentor Podcast · 21 min
Watch on YouTube →Rob Dial examines how unconscious values and societal programming lead to psychological entrapment and a sense of worthlessness. He argues that true freedom comes from consciously choosing values rather than chasing infinite metrics like wealth or validation.
- Worth management is often mistaken for time management when individuals use productivity to validate their self-worth.
- Chasing infinite metrics like money or knowledge leads to a mental state of addiction where one can never have enough.
- A person's greatest strength can become a liability when they lose the ability to stop or pivot to other life domains.
- Physical beauty and intellect are unstable foundations for identity because they inevitably fade or are surpassed by others.
- The rat race is defined as a mental state of unconscious worship rather than a specific career path or physical location.
The #1 AI Agent is Completely Free
EJazz · Limitless Podcast · 33 min
Watch on YouTube →This video explores the transition from conversational AI to autonomous agents that perform actions through virtual machines and browser navigation. EJazz discusses the competitive landscape between Meta, Grok, and OpenAI while introducing privacy models for agentic workflows.
- Meta Muse leads the consumer market with two million daily active users and a weekly free tier of 100 million tokens.
- The Hotel Room and Safe privacy model ensures agents operate in isolated and encrypted environments to protect user data.
- Proactive AI agents that initiate contact are more effective for the majority of users who struggle with prompt engineering.
- Startups like Instinct are pivoting from subscription models to taking a percentage of facilitated e-commerce transactions.
- GrokBot utilizes a shadowing technique to learn by observing user workflows before attempting to automate them autonomously.
- OpenAI Dots targets the enterprise sector by connecting to over 4,000 different tools and applications.
Rapidly Deploying Useful General-Purpose Robots at Walden Robotics
Walden Robotics · Weights & Biases · 40 min
Watch on YouTube →Walden Robotics presents a deployment-first philosophy that prioritizes putting general-purpose robots into production environments early to gather real-world data. This approach treats autonomy as a sliding scale that improves through iterative feedback loops rather than a binary state achieved in a lab.
- The deployment-first loop focuses on scaling the interaction between robot execution and human intervention in real-world settings.
- Large Behavior Models allow robots to navigate the unpredictable messiness of production environments more effectively than traditional pipelines.
- Immediate deployment for useful work generates higher-value training data than simulations or controlled laboratory environments.
- Robotics ROI is shifted from long-term research and development to immediate operational value through data flywheel effects.
- Market dominance in the robotics sector will likely be won by companies that scale their data collection loops most efficiently.
Building Blazing Fast AI-Native Apps: A Developer's Guide to the Database Underneath
Natty · Weights & Biases · 29 min
Watch on YouTube →Natty explains how AI-native applications require specialized database architectures to handle high-concurrency machine queries and low-latency vector searches. The session demonstrates using ClickHouse as a unified data layer to manage the massive volume of tool calls and reasoning traces generated by agents.
- Wall Shadow is an open-source tool that enables sub-second latency replication from Postgres to ClickHouse at 300,000 records per second.
- AI-native apps must optimize for machine speed because user retention drops significantly if response latency exceeds half a second.
- Integrating vector search and OLAP capabilities into a single system reduces the complexity of bolting on secondary database systems.
- Tracing and observability are critical for debugging why agents fail to follow expected reasoning paths during complex tasks.
- Internal agents with access to data warehouses can democratize data inspection for non-technical staff within an organization.
AI Scientists Are Here: Autonomous Labs and Synthesis Superintelligence: Periodic Labs
Liam Fedus · Latent Space · 85 min
Watch on YouTube →Liam Fedus discusses the concept of Synthesis Superintelligence, where AI models are integrated with physical laboratory equipment to accelerate scientific discovery. Periodic Labs aims to move beyond digital data by using high-throughput experimentation to ground AI reasoning in physical reality.
- The Matter Compiler system reverse-engineers synthesis steps for materials based on target properties to compress decades of trial and error.
- Periodic Labs targets a 140 IQ level for embedded AI in lab equipment to enable real-time material characterization.
- Capturing the process of science, including failed experiments and negative results, provides more valuable training data than final successes alone.
- Vertical integration of labs and models is necessary because existing public scientific data is heavily biased toward positive outcomes.
- High-throughput automation increases the surface area for luck, allowing AI to discover emergent behaviors like superconductivity.
Why Bigger Context Windows Won't Save Your Agent: Elizabeth Fuentes Leone, AWS
Elizabeth Fuentes Leone · AI Engineer · 15 min
Watch on YouTube →Elizabeth Fuentes Leone challenges the reliance on large context windows, arguing that they often lead to degraded reasoning and hallucinations. She introduces the Strands Agents framework to demonstrate how context engineering can improve agent reliability and reduce operational costs.
- The Lost in the Middle phenomenon causes models to ignore instructions placed in the center of large context windows.
- Context engineering involves externalizing data through memory pointers rather than stuffing raw logs or datasets into the prompt.
- The ESCI framework provides strategies to select, compress, and isolate information to maintain high model performance.
- Implementing a 50 percent summary ratio for historical messages helps keep conversation context relevant without overwhelming the model.
- Capping tool-call counts and using sliding-window managers prevents agents from entering infinite loops or reasoning spirals.
Securing AI at Scale: CoreWeave and CrowdStrike on What Comes Next
Jim Higgins · Weights & Biases · 37 min
Watch on YouTube →Jim Higgins and Bartley Richardson discuss the evolution of security as AI moves from experimental labs to mission critical enterprise infrastructure. They argue that security must be integrated into high performance compute fabrics to enable speed without compromising data integrity.
- Organizations must transition from human in the loop to human on the loop models to counter AI driven threats that exceed human reaction speeds.
- Security functions as a competitive advantage by providing the necessary guardrails for companies to move fast with sensitive data.
- CoreWeave scales workloads to thousands of NVIDIA H100 GPUs while maintaining a zero trust posture across the infrastructure.
- The cloud provider now represents the new perimeter for corporate intellectual property requiring deeper evaluation of specialized AI cloud security stacks.
- Proactive defense strategies are essential for protecting the massive data pipelines and GPU clusters used for training large language models.
- A shift toward autonomous defense models allows security to act as a foundational enabler rather than a traditional speed bump in the development cycle.
References
PeopleViren Baraiya (@virenbaraiya) · Kamalakannan Nandagopal · Pray · William Lyon (https://lyonwj.com) · Zach · Brian Armstrong · Kristoff · Mike · Martin · Matt Gibiec · Bernd · Alon Horev · Abhi Arya (x.com/abhiarya) · Hannah · Harsh · Andy Burnham · Boris Johnson · Tucker Carlson · Jacob Rees-Mogg · Jason Arday · Razmig Hovaghimian · Brian Sandon · Garrett Galow (x.com/g2theg) · Aaron Levie · Chuck Robbins · Sam Altman · Rick Devenuti · Jensen Huang · Matt Garman (x.com/mattsgarman) · Raghu Raghuram (x.com/RaghuRaghuram) · Andy Jassy · Jake Broekhuizen · Sarat Krishnan · Jacob · Elissa Epel (x.com/Dr_Epel) · Martin Picard · Wendy Mendes · Shri Koparkar · Eduardo · Hannah Coutan · Rob Dial (coachwithrob.com) · David Foster Wallace · Bob Marley · Mark Zuckerberg · Elon Musk · Noah Shin · Natty · Liam Fedus · Ekin Dogus Cubuk · Dima Bahdanau · Paul Dirac · Phil Anderson · Elizabeth Fuentes Leone (x.com/lizfue) · Jim Higgins · Bartley Richardson
ToolsNetflix Conductor · Orkes · Postman · Neo4j · Neo4j Agent Memory Service (NAMS) · Cline · Terminal Bench · SWE-bench · Skills Bench · agentskills.io · vLLM · CoreWeave · Kubernetes · Terraform · Bluebox (dynatrace.com) · VAST Data · LM Cache · MCP (Model Context Protocol) · ClickBench · Perplexity · Okta SSO (okta.com) · Doppler (doppler.com) · Amazon Bedrock (aws.amazon.com/bedrock) · AWS Graviton · AWS Trainium · NVIDIA GPUs · Claude (anthropic.com) · Lovable (lovable.dev) · WOW CLI · Atlas Platform · LangChain (langchain.com) · LangSmith · Context Hub · NVIDIA GB200 · NVIDIA GB300 · Vera Rubin · Redfish APIs · Raki · NVLink · Coolant Distribution Units · Vera CPU · Meta Muse · GrokBot · Instinct · OpenAI Dots · Reflexion · ClickHouse (clickhouse.com) · Wall Shadow · Postgres · Strands Agents · Matter Compiler · CrowdStrike · NVIDIA H100 GPUs