Build Wiz AI Show

Old prompt tricks like think step by step can now make Claude worse and pricier. In 16 minutes, Bob and Alice unpack Lance Martin's AI Engineer talk on getting more from frontier models.In this episode:- Prompt audits and the prompt cache: up to 72% cheaper- Evals Claude can climb, and hill climbing without overfitting- Days-long agents with Claude Managed AgentsSources: Lance Martin; Thomas Kwa et al.; Hamel Husain; Ethan MollickBig AI ideas, made clear for builders. luongnv.com/podcast

Greg Isenberg's X article makes the case for AI roll-ups: buy an established services business and run its back office with AI agents.His playbook: buy the customers and trust, build a rulebook and review checkpoints, and move margins from 5-10% toward 30-40%.Guests add balance: even Thrive needed more than off-the-shelf AI, and AI users get less accurate when they stop checking.Based on $5T opportunity: AI Roll Ups by GREG ISENBERG (@gregisenberg) · x.com$5T opportunity: AI Roll Ups by GREG ISENBERG (@gregisenberg) · x.comThe Revolution in Services: General Catalyst's AI Rollup Strategy, with Marc Bhargava by Nathan Labenz · cognitiverevolution.aiOpenAI hopes to automate IT and accounting with Thrive Holdings by Nicole Kobie · itpro.comCentaurs and Cyborgs on the Jagged Frontier by Ethan Mollick · oneusefulthing.orgGuests: Theo (presents The Revolution in Services: General Catalyst's AI Rollup Strategy, with Marc Bhargava), Maya (presents Centaurs and Cyborgs on the Jagged Frontier), Sam (presents OpenAI hopes to automate IT and accounting with Thrive Holdings)

Greg Isenberg says the marketing engineer, part marketer and part builder, is the new 10x employee. Bob and Alice unpack his playbook: write the business "brain" before any agent, connect your data, add agents one at a time, and keep the guardrails in the tools. Guests Lena, Theo and Sam bring in Sean Ellis on growth hacking, Meta's engineering post on Andromeda and NPR on labelling AI content. Plus Builder's Corner: one thing to try this week. Every claim is checked against its source. Sources are credited in the episode

New here? Meet Bob and Alice, who swap roles every episode, and our four recurring guests: Maya, Lena, Theo and Sam. We explain how an episode is made: each source read in full, tested against outside evidence, and every claim checked word for word. Voices are AI; sources are credited in every episode. Big AI ideas, made clear for builders.

AI agents keep getting better, so the new limit is your attention: most people can follow only three to five agent sessions at once. Nora and Ben explain the DeepSeek Harness team's design for a "super-agent", one agent that is always on and works on many tasks at once. A front-desk router talks to you and hands out tasks to background workers. Cheap code filters out the noise, and the agent can quietly fix bugs before anyone asks. The hosts also go through the team's early test results and the main unsolved problem, security.Source: DeepSeek Harness on X, "Exploring Super-agent Design" (1 Nov 2026).

Pi Durable is an experimental TypeScript agent harness designed to create resilient, long-running agent applications that reliably survive crashes and process interruptions. By continuously committing model turns, tool calls, and application states to persistent storage backends like SQLite or JSONL, the library ensures that interrupted tasks resume seamlessly from their exact checkpoints without duplicating side effects. Unlike traditional terminal-bound coding assistants, this foundational building block enables developers to construct proactive agents that initiate tasks from external events, support multi-client interactions, and allow human oversight near the loop. Furthermore, it integrates smoothly with diverse JavaScript runtimes and model providers, offering advanced capabilities such as background compaction, safe replay policies, and multiplayer workflows.

Decision models represent a major shift in enterprise artificial intelligence by prioritizing rapid, structured choices over traditional text generation. Instead of relying on slow large language models to produce unstructured essays, these specialized systems process input data to deliver typed outputs like choices, scores, and probabilities in a single forward pass. Platforms such as TypeSafe's Jev and various open-source alternatives introduce unprecedented speed and cost-efficiency while eliminating formatting errors entirely. By providing calibrated confidence scores alongside every decision, these models enable software architectures to gauge their own uncertainty effectively. Consequently, organizations can seamlessly combine decision models with conventional language models to establish a secure control plane for automated agents and high-frequency workflows.

Dario Amodei advocates for a strategic deceleration of artificial intelligence development to ensure that safety measures can keep pace with rapidly advancing capabilities. He highlights the dual nature of AI, acknowledging its potential to eradicate diseases while warning against catastrophic risks like bioterrorism and autonomous system failures. Amodei introduces a "pacing" framework that begins with embedded third-party evaluators to verify safety protocols within companies like Anthropic. His proposal extends to democratic and global coordination, aiming to establish common standards while maintaining a technological lead over authoritarian regimes. Ultimately, the author argues that prioritizing prudence over speed is essential to preventing misalignment and securing a beneficial future for humanity.

The concept of an AI agent harness, using the analogy of a climbing harness to describe its protective and functional role. Just as a physical harness secures a climber and organizes their gear, a software harness provides a controlled environment where AI models can operate safely and effectively. The source identifies four primary components: a system prompt for guidance, a set of digital tools, an agentic loop for autonomous task completion, and a translation layer for model flexibility. By utilizing an open-source harness, users gain greater agency and the freedom to switch between different AI providers while maintaining ownership of their data. Ultimately, the author argues that these frameworks are essential for ensuring that humans remain in control of artificial intelligence rather than being dominated by it.

@skills, an open protocol designed to revolutionize how AI agents access and manage procedural knowledge. Current systems rely on a restrictive installation model that forces skill descriptions into a permanent "resident" prompt, leading to token waste, decreased model attention, and a "trigger lottery" where skills often fail to activate. To solve this, @skills separates the content of a skill from its persistence and its ability to auto-trigger, offering a three-tier delivery system that ranges from on-demand references to permanent residency. This protocol utilizes a simple @syntax and a file-tree architecture, allowing users to invoke thousands of skills by name without overtaxing the AI's limited attention budget. Furthermore, it integrates with a central skills hub for discovery while maintaining local-first resolution and compatibility with existing agent workflows. Ultimately, the protocol aims to make the vast "long tail" of specialized AI skills reachable, manageable, and effective for both individual developers and collaborative teams.