Sunday, October 4, 2026
The agent is the new user.
October 4 · 9 videos
OpenAI is launching Dots.
Agents will soon outnumber humans online.
Zuckerberg is the last man standing.
Dashboards are dead.
The human is now an async API.
AssemblyAI resolved 80 percent of tickets with Joey.
Sourcegraph warns of AI codebase decay.
“The software that runs the world is not pretty. It's not new. It's not clean. It's how everything actually works.”
Agents That Write Their Own Tools at Runtime — Sandhya Subramani, AWS
Sandhya Subramani · AI Engineer · 20 min
Watch on YouTube →Sandhya Subramani introduces the Strands Agents SDK, which allows AI to create and execute its own tools at runtime. This shift moves agents from static toolsets to self-evolving systems.
- Agents start with a base of three tools: editor, shell, and load_tool.
- The system prompt defines what a good tool looks like to guide autonomous creation.
- Dynamic agents can write Python code to solve missing capabilities without a restart.
- Framework allows orchestration of complex workflows using Swarm and Graph patterns.
- Subramani emphasizes an 8-point evaluation hierarchy for safe production deployment.
- A Python version of the agent successfully wrote its own TypeScript port port independently.
AI Coding Agents Are Breaking Big Codebases — Dan Adler, Sourcegraph
Dan Adler · AI Engineer · 12 min
Watch on YouTube →Dan Adler, CEO of Sourcegraph, warns that AI coding agents are causing large-scale codebase decay in critical enterprise systems. He argues that visibility is the primary bottleneck for agentic maintenance.
- 72 percent of software industry employment is in companies with over 500 people.
- Agents cannot effectively maintain codebases they cannot see or search at scale.
- Adler introduces Agentic Batch Changes to execute and audit code across thousands of repos.
- The Code Graph provides a compiler-accurate map for agents to understand historical context.
- Raw code generation increases the cost of ownership by creating more code to secure.
- Confidence in AI is currently undermined by the problem of not knowing what generated code does.
Keyword Search Is Dying. Is Your Catalog Ready for AI Agents? — PayPal
Nixon Dinh · AI Engineer · 16 min
Watch on YouTube →Nixon Dinh of PayPal discusses the transition from the Search Era to the Intent and Delegation Eras. He argues that product catalogs must be optimized for AI agents rather than human SEO.
- Agents are projected to execute 15 to 25 percent of all commerce by 2030.
- AI agents will eventually outnumber humans online within a single decade.
- The agentic commerce opportunity in US retail is estimated at 1.1 trillion dollars.
- Conversion rates are 4 times higher when users interact with an agent before ordering.
- Over-enriching catalogs with unstructured boilerplate leads to agent hallucinations.
- Brands must prioritize structured content quality and trust signals over brand recognition.
The Human Is an Async API — Melanie Warrick, Temporal
Melanie Warrick · AI Engineer · 19 min
Watch on YouTube →Melanie Warrick explains why humans should be treated as asynchronous APIs in agent architectures. She advocates for durable execution to handle long-running human-in-the-loop tasks.
- Human response times range from minutes to weeks, requiring durable state preservation.
- Temporal primitives: Workers, Workflows, and Activities: ensure deterministic orchestration.
- State remains consistent even if worker processes are killed during a wait for approval.
- Developers should assess the cost of being wrong to determine human intervention needs.
- Reliability is the prerequisite for moving from experimental agents to production systems.
- The framework allows scaling to millions of concurrent, long-running processes.
We Built an AI Support Agent That Resolves 80% of Tickets — AssemblyAI
Matt Lawler · AI Engineer · 16 min
Watch on YouTube →Matt Lawler details how AssemblyAI built Joey, an in-house support agent that handles 80 percent of tickets. The project demonstrates the value of forward-deployed engineers automating their own roles.
- Joey increased ticket resolution from 10 percent to 80 percent compared to vendor bots.
- The agent uses 30,000 lines of markdown instructions and local documentation embeddings.
- Running the custom solution on Railway costs only 700 dollars per month.
- Human escalations are treated as the prioritized roadmap for the agent's next features.
- Joey can perform complex tasks like pricing negotiations and legal agreement guidance.
- AssemblyAI dogfooded its own Voice Agent API to give Joey a real-time voice interface.
OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux
Tibo Sottiaux · Lenny's Podcast · 37 min
Watch on YouTube →Tibo Sottiaux, Head of ChatGPT, introduces Dots, a new personal agent platform. He predicts a shift toward permanent active intelligence where agents learn user goals 24/7.
- OpenAI is moving away from manual user-managed loops toward invisible AI extensions.
- The most valuable skill in the agentic era is great taste and user empathy.
- OpenAI is implementing a revenue share model for popular developer plugins.
- A founder mindset is the most successful archetype for high-growth AI organizations.
- New graduates have an advantage by learning to work natively with AI from the start.
- 120 former YC founders are currently working at OpenAI to drive bottom-up energy.
How Mark Zuckerberg, Palmer Luckey, Josh Kushner & Scott Wu Think | Jeremy Stern
Jeremy Stern · David Senra · 68 min
Watch on YouTube →Jeremy Stern analyzes the psychological blueprints of elite founders like Mark Zuckerberg. He explores the competitive drive and sense of duty that fuels high-level achievement.
- Zuckerberg is described as the last man standing of his generation of founders.
- The Obligation to the Gift: failing to exploit unique talent is viewed as a personal sin.
- Meta has transitioned from a growth-focused pillage state to a stable trade empire.
- 50 percent of the human population uses Meta apps like WhatsApp and Instagram.
- Stripe is cited as potentially the second most valuable millennial-founded company.
- Stern applies literary techniques to reveal the human realities of tech leaders.
Dashboards Are Dead — Sarah Simionescu, Composio
Sarah Simionescu · AI Engineer · 10 min
Watch on YouTube →Sarah Simionescu argues that software interfaces must shift from human-centric dashboards to agent-centric execution layers. She highlights the inefficiencies of current visual translation layers.
- AI agents are becoming the primary users of software, rendering dashboards obsolete.
- The Model Context Protocol (MCP) brings agents closer to data but lacks memory.
- Agents should not have to relearn formatting or navigation in every new session.
- Composio uses tool search and execution plans to fix bugs in under five minutes.
- GitHub's toolkit in Composio contains over 200 specialized tools.
- Competitive advantage will be defined by how judgeable a product is by an agent.
Stop Fine-Tuning to Fix Retrieval Problems — Anant Srivastava
Anant Srivastava · AI Engineer · 20 min
Watch on YouTube →Anant Srivastava dismantles the idea that fine-tuning is an escalation of RAG. He provides a framework for deciding where knowledge should live: prompts, memory, or weights.
- Prompts manage small, stable behaviors; memory handles large, current facts.
- Weights should only be used for fixed reflexes or reasoning patterns that have stabilized.
- Fine-tuning to solve retrieval problems often leads to hallucinations of old data.
- The Circulating Architecture moves stable retrieval patterns into model weights over time.
- Enterprise systems must maintain citability, which is lost when facts are baked into weights.
- The model is the easy part: the harness that routes knowledge is the true advantage.
References
PeopleSandhya Subramani · Dan Adler (https://x.com/DanielNealAdler) · Nixon Dinh · Melanie Warrick (https://x.com/nyghtowl) · Matt Lawler · Tibo Sottiaux · Ahmed Ibrahim · Peter Steinberger (@steipete) · Sam Altman · Jeremy Stern · Mark Zuckerberg · Palmer Luckey · Scott Wu · Josh Kushner · Rick Rubin · Sarah Simionescu (https://www.sarahsimionescu.com/) · Anant Srivastava
ToolsStrands Agents SDK · Sourcegraph · Agentic Batch Changes · Code Graph · PayPal · Temporal · Google ADK · LangGraph · AssemblyAI · Joey · Claude Agent SDK · Voyage · Railway · ChatGPT · Dots · Meta · WhatsApp · Instagram · Stripe · Composio · Model Context Protocol (MCP) · GitHub · Slack · DataDog