Nan Jiang introduces "Adam absorption" to revolutionize RL model synchronization. By exploiting finite precision serving and small Adam steps, less than 1% of served model weights actually change, allowing for 500MB patches instead of 500GB checkpoints. This enables a distributed "bulletin board" architecture, decoupling trainers from global rollout fleets and unlocking elastic, cross-region GPU capacity for RL.
Once GPT 5.3 Codex Spark achieved 1000 tokens/sec on Cerebras, network latency superseded inference as the bottleneck for agents. This talk details how the Codex harness addresses this and other agentic challenges through innovations like WebSocket mode for stateful context, deferred tools for efficient context construction, robust sandboxing (Seatbelt, Bubblewrap, custom Windows solution), and an auto-review subagent to mitigate approval fatigue while ensuring security. It also covers structured actions via 'apply patch' for file edits, shell tools for system interaction, and sophisticated long-horizon goal management, with most distinct features exposed through the open Responses API.
Alejandro Maza Ayala, Chief Product & AI Officer at Kavak, details how the used-car marketplace transformed into an AI-native company. He explains the 'agent-per-customer' architecture, where individual agents handle 96% of customer interactions and 95% of transactions, outperforming human teams in sales (2.1x better conversion) and even acting as an 'AI CEO' that boosted profits by 50% in an experimental city. The discussion covers the need to redesign company structures, the importance of robust evaluations, and how a 'Jedi Academy' trains all employees, from executives to mechanics, to build and collaborate with AI agents. Ayala argues for 'creative destruction,' suggesting that true AI leverage comes from rebuilding organizations from the ground up, rather than incremental adoption, presenting a massive opportunity for new founders.
This video outlines five essential best practices for developing reliable, secure, and effective AI agent skills. It covers optimizing skill triggering through descriptive metadata, leveraging real-world domain expertise over generic LLM output, managing context windows efficiently by writing lean skills and using progressive disclosure, implementing deterministic logic with scripts for fragile operations, and critically vetting all skills for security vulnerabilities before deployment. These practices are crucial for professionals building robust agentic systems.
Arjun Singh, co-founder of Superconductor, discusses the six crucial lessons learned from integrating AI agents into their software development workflow. He emphasizes the importance of building "multiplayer agentic engineering" systems where human teams and AI agents collaborate seamlessly, focusing on model agnosticism, pervasive agent interfaces, transparent work visibility, automated signal-to-code conversion (like their "meeting bot"), secure isolated cloud environments, and internal code-base benchmarking to optimize cost, quality, and speed.
Anirban Chatterjee discusses the emergence of "verification debt" in AI-driven software development, where AI-generated code introduces persistent quality issues despite initial productivity gains. He proposes a **zero-trust, multi-layered verification** approach, integrating it into an "Agent-Centric Development Cycle" (ACDC) to ensure code quality, security, and compliance by empowering AI agents with real-time feedback and remediation capabilities through tools like SonarQube, Gitarr, and Sonar Vortex.