Thursday, May 21, 2026
The era of the autonomous agentic enterprise has arrived.
May 21 · 27 videos
Cerebras hit a $63B IPO.
OpenAI signed a $20B deal for wafer-scale chips.
Daytona is spinning up 850,000 sandboxes daily.
Erik Thorelli says code review is the new bottleneck.
Barun Singh predicts a debt realization event in 2026.
YC startups hit 5x revenue per employee.
Atai Barkai says all UI will be AI.
“Routing is the architectural conclusion of evaluation.”
AI Dev 26 x SF | Ashwyn Sharma: Every App Needs a Voice UI. Here's How to Build It
Ashwyn Sharma · DeepLearningAI · 16 min
Watch on YouTube →Vocal Bridge CEO Ashwyn Sharma explains why Voice UI is becoming an essential interface for every application. He details how to move past the technical hurdles of voice activity detection and turn detection.
- Voice interfaces are moving from nice-to-have to an industry standard like mobile apps.
- Brainstorming is more natural via voice than text, making it a high-value entry point.
- Building production voice AI from scratch typically takes six months to years.
- Vocal Bridge abstracts voice complexity into three core surfaces for developers.
- A technical framework allows bi-directional communication between React UI and voice agents.
- Decoupling the dialogue brain from the LLM prevents context window bloat.
AI Dev 26 x SF | Atai Barkai: Fullstack Agents & Generative UI with AG UI
Atai Barkai · DeepLearningAI · 17 min
Watch on YouTube →Atai Barkai of CopilotKit argues that all UI will eventually be AI-mediated. He introduces the AG UI protocol to manage the transition from chatbots to integrated agentic interfaces.
- AI is becoming a mediated layer between humans and all technology, including appliances.
- The AG UI protocol manages the spectrum from deterministic components to open-ended code.
- Data labeling is shifting from external low-wage work to internal expert-driven feedback.
- Continuous learning loops where agents observe human edits provide a massive competitive advantage.
- Mass adoption requires moving past text boxes to intuitive, GUI-like agentic experiences.
- CopilotKit powers tens of millions of interactions every week for Fortune 500 companies.
AI Dev 26 x SF | Idan Raman: The Identity Crisis of Browser Agents
Idan Raman · DeepLearningAI · 16 min
Watch on YouTube →Idan Goldman, CEO of Anchor, discusses the security bottleneck of adapting human-centric web security for AI agents. He warns against the build vs. buy trap for non-core infrastructure.
- Adapting 20 years of human identity technology for agents is a massive enterprise bottleneck.
- Agent identity involves complex bot detection, network reputation, and MFA management.
- Stage Zero infrastructure like specialized agent VPNs is required for reliable deployment.
- The Complexity Creep model shows how simple projects can derail into deep rabbit holes.
- Passing credentials in prompts is insufficient as agents can leak data after parsing pages.
- Authentication must be viewed as a continuous managed lifecycle rather than a transaction.
AI Dev 26 x SF | Tom Howlett: Can LLMs Generate Enterprise Quality Code?
Tom Howlett · DeepLearningAI · 36 min
Watch on YouTube →Tom Howlett from Sonar addresses the gap between initial AI coding velocity and long-term technical debt. He introduces the Agent-Centric Development Cycle to maintain code quality.
- Initial 3-5x velocity gains from agents often vanish within three months due to technical debt.
- Sonar's benchmark of 52 models reveals massive variance in code quality and security bugs.
- The ACDC framework shifts static analysis from the outer loop to the inner IDE loop.
- Vibe coding works for small tools but fails for mission-critical production applications.
- Automated verification is necessary because human review takes 95 percent of the time.
- Some models produce twice as much code as others for the exact same functional task.
AI Dev 26 x SF | Erik Thorelli: Deploying AI Code Review at Scale
Erik Thorelli · DeepLearningAI · 30 min
Watch on YouTube →Erik Thorelli of CodeRabbit explains how AI has shifted the software bottleneck from generation to review. He advocates for an evals-first culture to manage the increase in bugs.
- AI-generated code results in a 40 percent increase in critical bugs compared to human code.
- Context engineering must move beyond simple RAG toward deterministic repository enrichment.
- Routing requests across model families is the architectural conclusion of rigorous evaluation.
- Internal benchmarks are more valuable than public ones like SWE-bench for product context.
- The cost of downtime is estimated at 5 million dollars per hour for major enterprises.
- Continuous deployment allows for easier bisection of regressions in probabilistic systems.
AI Dev 26 x SF | Brandon Middleton: Vibe Coding Master Class
Brandon Middleton · DeepLearningAI · 32 min
Watch on YouTube →Brandon Middleton from Replit argues that the developer paradigm has shifted from syntax to natural language prompts. He outlines a new framework for job readiness in 2026.
- The unit of work has shifted from a line of code to a natural language specification.
- Vibe coding prioritizes clear specification and system orchestration over manual implementation.
- AI is subsuming entry-level work, requiring new graduates to have the judgment of 3-year pros.
- Job readiness now requires the ability to prototype vague problems in under 24 hours.
- Education should shift from abstract hackathons to real-world build-a-thons.
- An eternal student mindset is mandatory as the delta of change occurs on a weekly basis.
AI Dev 26 x SF | Ondra Urban: Agents with Wallets? Putting 25,000 Tools on x402
Ondra Urban · DeepLearningAI · 21 min
Watch on YouTube →Ondra Urban from Apify discusses the necessity of agentic payment protocols for true autonomy. He highlights the x402 protocol as a solution for decentralized, USDC-based transactions.
- Agents must be able to discover and purchase resources without human intervention.
- The current paradigm of hardcoded API keys is unscalable and insecure for agents.
- Coinbase's x402 protocol allows implementation via middleware without altering API bodies.
- Apify implemented an exact scheme with refunds to handle usage-based pricing for tools.
- The MCPC CLI solves security risks by keeping wallet credentials in the system keychain.
- Apify's developer community earned 1.2 million dollars in payouts last month.
AI Dev 26 x SF | Aman Singla & Aseem Chandra: MarcoPolo, A Workspace for AI to Work with Your Data
Aman Singla · DeepLearningAI · 18 min
Watch on YouTube →Aman Singla and Aseem Chandra introduce MarcoPolo, a secure middleware for agentic data operations. They argue that agents need a persistent workspace to perform complex data tasks.
- Agents lack a dedicated environment with local compute to perform cross-system joins.
- MarcoPolo provides a Kubernetes-based container with DuckDB and a unified data CLI.
- Enterprise AI adoption is gated more by security anxiety than by model capability.
- Treating AI integration like onboarding a new employee is a powerful mental model.
- Schema-on-connect solves the cold start problem by curating documentation for the agent.
- The system transforms fragmented interactions into a compounding organizational knowledge base.
AI Dev 26 x SF | Barun Singh & Kennith Jackson; The Hidden Cost of AI Velocity and AI Agents
Barun Singh · DeepLearningAI · 34 min
Watch on YouTube →Barun Singh predicts a debt realization event in late 2026 caused by unvetted AI-generated code. He advocates for supervised agents over fully autonomous ones for production.
- Short-term gains in PR velocity are masking a looming crisis of technical debt.
- A debt realization event is expected in late 2026 as teams realize they sacrificed reasoning for speed.
- Software engineering remains an iterative craft that requires human modeling and communication.
- Speeding up coding often just exposes bottlenecks in QA and organizational bureaucracy.
- The market for specific AI engineering tasks like RAG is rapidly commoditizing.
- Expertise is defined by the ability to communicate technical work at multiple levels of abstraction.
AI Dev 26 x SF | Matthew Xu: The 4-Legged Identity Challenge
Matthew Xu · DeepLearningAI · 16 min
Watch on YouTube →Matthew Xu discusses the security challenges of scaling Model Context Protocol systems for enterprise use. He introduces the 4-legged identity problem for agentic delegation.
- Standard OAuth patterns are not designed to handle agents acting as intermediaries.
- The transition from local to shared AI infrastructure mirrors the shift from desktop to SaaS.
- Least-privilege access and privacy-aware logging are prerequisites for enterprise trust.
- Weak auditing is a primary failure mode that prevents agents from meeting compliance standards.
- Solving identity is a prerequisite for moving agents from experiments into production.
- Developer-friendly security patterns are the only way to ensure adoption in fast-moving teams.
AI Dev 26 x SF | David Park: Building Production Grade Agentic Systems with ADE
David Park · DeepLearningAI · 29 min
Watch on YouTube →David Park of LandingAI presents a blueprint for production-grade agents in regulated industries. He focuses on Agentic Document Extraction as the foundation for decision-making.
- Documents are the primary source of truth for most real-world enterprise decisions.
- Production AI requires a deterministic shell around a stochastic model core.
- A global bank achieved a 60 percent reduction in manual review time using this architecture.
- Auditability and explainability are non-negotiable in financial services and healthcare.
- Hierarchical agent architectures separate data ingestion from logical reasoning.
- Schemas should be defined before prompts to ensure robust context engineering.
AI Dev 26 x SF | Tushar Jain: Shipping Agents Safely, Boundaries That Actually Work
Tushar Jain · DeepLearningAI · 30 min
Watch on YouTube →Tushar Jain from Docker argues that prompt guardrails are insufficient for agents with write access. He introduces SPX as a secure runtime for agent containment.
- Prompt guardrails shape intent but cannot enforce it once an agent has write access.
- Docker's SPX provides a distributed secure runtime using isolated micro-VMs.
- Safety is the prerequisite for the autonomy required to run agents continuously.
- The run yellow mindset allows for speed within strictly governed safe boundaries.
- Credential injection must happen outside the execution environment to prevent leaks.
- Coding agents crossed a chasm in late 2024 where they became highly addictive and capable.
AI Dev 26 x SF | Ankit Mathur: The Coding Agent Multiverse of Madness
Ankit Mathur · DeepLearningAI · 26 min
Watch on YouTube →Ankit Mathur of Databricks discusses the governance challenges of scaling coding agents. He advocates for a gateway architecture to manage security and costs.
- Coding agent usage is scaling 10 to 100 times faster than all other AI agents combined.
- Enterprises face a multiverse of madness with fragmented tools and security liabilities.
- The AI Gateway provides centralized observability while allowing developer tool freedom.
- Databricks uses an internal tool called Isaac to manage 10,000 employees using agents.
- Organizational velocity is limited by the ability to automate performance validation.
- Market norms have shifted to require 10x AI-augmented engineers for innovation.
AI Dev 26 x SF | Amrita Venkatraman: 3rd Era of Software Development
Amrita Venkatraman · DeepLearningAI · 31 min
Watch on YouTube →Amrita Venkatraman describes the shift to a third era of software development defined by autonomous agent systems. She notes that 60 percent of enterprise code is already AI-touched.
- The third era moves beyond autocomplete to fleets of cloud-based agents on remote VMs.
- Over 60 percent of code lines in enterprise products were touched by AI as of March 2024.
- Engineering constraints are shifting from technical capacity to the quality of ideas.
- An agent system built a functional browser from scratch in one week with 3 million lines of code.
- Humans are shifting from translation layers to system architects and agent managers.
- Autonomous systems can handle toil like feature flag cleanup and incident post-mortems.
AI Agents Need Computers: 74% MoM Growth, 850K/Day Runs, & New Agent Cloud — Ivan Burazin, Daytona
Ivan Burazin · Latent Space · 71 min
Watch on YouTube →Daytona CEO Ivan Burazin outlines the shift to composable computers for AI agents. He reports explosive growth in the demand for stateful agent sandboxes.
- Daytona is seeing 74 percent month-over-month growth driven by agent compute needs.
- One customer runs approximately 850,000 sandboxes daily for agentic workloads.
- The architecture enables a 60ms startup time for environments without Kubernetes overhead.
- Agents require full access to legacy Windows and macOS applications to act as human emulators.
- The bottleneck is shifting from GPU availability to CPU and memory for running agent fleets.
- AI infrastructure will eventually resemble Stripe's API-first, consumption-based model.
Relational Foundation Models for Enterprise Data [Jure Leskovec] - 768
Jure Leskovec · The TWIML AI Podcast with Sam Charrington · 65 min
Watch on YouTube →Stanford Professor Jure Leskovec argues that manual feature engineering is a bottleneck for enterprise ML. He introduces Relational Deep Learning as a superior alternative.
- Enterprise data is best represented as graphs of interacting entities rather than single tables.
- Manual feature engineering introduces human bias and ignores rich relational structures.
- The RFM2 foundation model can reason over any structured relational data via in-context learning.
- Relational models outperform traditional supervised models by 5 to 12 percent out-of-the-box.
- Reddit and DoorDash have seen double-digit increases in metrics like click-through rates.
- ML progress is moving from CPU-based summaries to GPU-based raw data reasoning.
Has China Trump-ed The USA?
Alastair Campbell · The Rest Is Politics · 40 min
Watch on YouTube →Alastair Campbell and Rory Stewart analyze the geopolitical shifts following the Trump-Xi summit. They discuss the retreat of Western leadership and the rise of demographic collapse.
- Xi Jinping asserts that a century-long transformation of global power is accelerating.
- Drastic cuts to UK and US international development are creating dangerous power vacuums.
- Over half of Reform UK voters support the forced removal of non-white citizens born abroad.
- Plummeting birth rates globally show a direct correlation with high-speed internet access.
- Polarization is driven by algorithms of division that fuel gender-based ideological splits.
- The world is vulnerable to a new Ebola outbreak in the DRC due to dismantled health funding.
Cooking with Agents in VS Code — Liam Hampton, Microsoft
Liam Hampton · AI Engineer · 17 min
Watch on YouTube →Liam Hampton presents a framework for orchestrating local, background, and cloud agents in VS Code. He argues against the one-shot prompt fallacy for developer productivity.
- Developer productivity comes from matching specific tasks to local, background, or cloud agents.
- Local agents handle high-context tasks while cloud agents handle low-touch overhead.
- VS Code is evolving into a unified control plane for multiple AI models and agents.
- The Model Context Protocol is used to extend agent capabilities within the IDE.
- Token expenditure can be optimized by instructing agents to use concise language.
- Reducing cognitive load requires a human-in-the-loop gradient based on task risk.
AI Dev 26 x SF | Vlad Luzin: Herding Cats—The Hidden Challenges of Multi-Agent Autonomy
Vlad Luzin · DeepLearningAI · 30 min
Watch on YouTube →Vlad Luzin discusses the evolution of multi-agent systems into autonomous distributed networks. He introduces the Banff platform as an AI Mesh for cross-framework interaction.
- Agents are like cats, not dogs, and will not follow rigid instructions perfectly.
- The future of business interaction is AI-to-AI communication using natural language as an API.
- Banff enables real-time interaction between agents built on LangGraph, CrewAI, and Pydantic AI.
- Developers currently waste time acting as manual message buses between different models.
- Enterprises require a governance layer for agent identity and multi-tenant audit trails.
- Autonomous capabilities for coding agents now range from ten minutes to a full day.
AI Dev 26 x SF | Melissa Herrera: Your Agents Should Be Durable
Melissa Herrera · DeepLearningAI · 27 min
Watch on YouTube →Melissa Herrera explains why durable execution is necessary for production-ready AI agents. She uses Temporal to ensure agents can resume from failures without losing state.
- AI agents in production face inevitable infrastructure failures and API rate limits.
- Durable execution ensures agents resume from the point of failure rather than restarting.
- OpenAI uses Temporal to orchestrate complex, long-running tasks like image generation.
- Naive restarts waste expensive tokens and risk non-deterministic outcomes.
- The goal is self-healing code that recovers from failures without human intervention.
- Production-readiness is a competitive advantage over smarter but fragile agents.
AI Dev 26 x SF | Carter Rabasa: File Systems Are the New Primitive for AI Agents
Carter Rabasa · DeepLearningAI · 32 min
Watch on YouTube →Carter Rabasa argues that file systems are the critical primitive for building robust AI agents. He explains why LLMs are natively proficient at navigating hierarchical structures.
- LLMs have an inherent intuition for file systems because their training data is mostly code.
- File systems provide a structured and durable way for agents to organize long-term memory.
- Using established technologies like file systems is often more effective than novel data structures.
- File-based workflows create a natural audit trail for human-in-the-loop collaboration.
- Standardizing on file systems accelerates development by using well-understood infrastructure.
- Directories are a universal interface for data interoperability between different agents.
Scaling Agents on Kubernetes with acpx and ACP — Onur Solmaz, OpenClaw
Onur Solmaz · AI Engineer · 19 min
Watch on YouTube →Onur Solmaz discusses managing the firehose of AI-generated pull requests using Kubernetes. He introduces acpx for structured agent-to-client communication.
- OpenClaw manages 300 to 500 AI-generated pull requests per day.
- The acpx CLI replaces brittle PTY scraping with structured Agent Client Protocol communication.
- Spritz is a Go operator that provisions disposable agent pods on Kubernetes.
- Discord Driven Development allows managing parallel agent sessions from a mobile device.
- AI-generated PRs provide crucial signal about where a codebase causes friction for users.
- Enterprise adoption requires automated provisioning of agent identities like Slack apps.
How to Build a Self-Improving Company with AI
Tom Blomfield · Y Combinator · 13 min
Watch on YouTube →Tom Blomfield argues that AI renders the traditional Roman Legion corporate hierarchy obsolete. He advocates for architecting companies as recursive, self-improving AI loops.
- Modern YC startups achieve 5x more revenue per employee than those from 18 months ago.
- The Copilot model is flawed because it grafts tech onto inefficient human processes.
- Companies should extract domain knowledge from Slack and emails to create a company brain.
- Middle management is eliminated in favor of individual builders and single DRIs.
- Software becomes ephemeral while the underlying business context becomes the primary value.
- Humans should focus on high-stakes emotional intelligence and novel real-world interactions.
When You Realize No One Cares... Everything Changes
Rob Dial · The Mindset Mentor Podcast · 16 min
Watch on YouTube →Rob Dial discusses the Spotlight Effect and how imagined social judgment limits personal freedom. He provides frameworks for breaking free from people-pleasing.
- The Spotlight Effect causes people to dramatically overestimate how much others notice them.
- Social rejection activates the same brain regions as physical pain due to tribal evolution.
- A Cornell study showed only 20 percent of peers notice flaws that participants thought 50 percent would.
- Social judgment is typically a projection of the judge's own internal wounds.
- Freedom is directly tied to the willingness to be embarrassed in pursuit of growth.
- High performance requires transitioning from self-monitoring to authentic presence.
Your Coding Agent Should Do AI System Engineering — Ben Burtenshaw, Hugging Face
Ben Burtenshaw · AI Engineer · 18 min
Watch on YouTube →Ben Burtenshaw of Hugging Face explains how coding agents are now tackling complex systems engineering. He details how agents can optimize CUDA kernels for massive speedups.
- Agents have crossed an acceptance gradient to handle optimized CUDA kernel writing.
- An agent-written RMSNorm kernel achieved a 1.88x speedup on H100 GPUs.
- AutoLab is a multi-agent research architecture that automates end-to-end research pipelines.
- Developers must move closer to the silicon as high-level coding becomes commoditized.
- The Hugging Face Hub is transitioning into an active compute and tracking layer for agents.
- Successful automation requires exposing primitives to agents rather than abstracting them.
Essentials: The Science of Learning & Speaking Languages | Dr. Eddie Chang
Eddie Chang · Andrew Huberman · 28 min
Watch on YouTube →Dr. Eddie Chang discusses the neurobiology of speech and the success of the BRAVO trial. He explores the future of brain-machine interfaces for restoring communication.
- Speaking is the most complex motor feat of the human species.
- The BRAVO trial restored communication to a paralyzed man using a 50-word neural decoder.
- Stuttering is a coordination failure of a neural symphony, not a language deficit.
- Auditory feedback is crucial for fluency as the brain constantly monitors produced sounds.
- Neurotechnology is transitioning from academic research into commercial medical products.
- Future digital avatars will incorporate non-verbal facial expressions for holistic communication.
The Story Behind Cerebras’ $63 Billion IPO with Founder and CEO Andrew Feldman
Andrew Feldman · No Priors: AI, Machine Learning, Tech, & Startups · 30 min
Watch on YouTube →Cerebras CEO Andrew Feldman discusses the company's 63 billion dollar IPO and the shift to wafer-scale computing. He details a landmark 20 billion dollar deal with OpenAI.
- Cerebras chips are the size of a dinner plate and achieve 20x faster inference than GPUs.
- The company survived a brutal phase burning 8 million dollars per month during R&D.
- A 20 billion dollar deal with OpenAI was executed in just four weeks in 2025.
- High-speed inference became a must-have once models were smart enough to be useful.
- The Professional David mindset involves using intellectual agility to outmaneuver incumbents.
- Ultra-fast AI will enable a fundamental reorganization of business models like broadband did.
References
PeopleAshwyn Sharma · Atai Barkai · Idan Raman · Tom Howlett · Erik Thorelli · Brandon Middleton · Ondra Urban · Aman Singla · Aseem Chandra · Barun Singh · Kennith Jackson · Matthew Xu · David Park · Tushar Jain · Ankit Mathur · Amrita Venkatraman · Ivan Burazin (x.com/ivanburazin) · Jure Leskovec · Alastair Campbell · Rory Stewart · Liam Hampton (x.com/liamchampton) · Vlad Luzin · Melissa Herrera · Carter Rabasa · Onur Solmaz · Tom Blomfield · Rob Dial · Ben Burtenshaw (x.com/ben_burtenshaw) · Eddie Chang · Andrew Feldman
ToolsVocal Bridge · CopilotKit · AG UI · Anchor · Sonar · CodeRabbit · Replit Agent 4 · Apify · x402 · MarcoPolo · DuckDB · Andela · Docker SPX · Databricks AI Gateway · Cursor · Daytona · Kumo.ai · VS Code · Banff · Temporal · OpenClaw · acpx · Cerebras WSE