Claude's new ability to learn workflows through screen recordings effectively kills the era of prompt engineering by shifting the competitive advantage to domain experts who can simply demonstrate their tacit knowledge to the model.
Only the stories worth your time.
Get the next daily digest delivered to your inbox — curated from trusted sources and summarized in minutes. No spam.
Editor's Notes
The shift toward agentic AI is moving away from complex prompt chains and toward direct, visual interaction with existing software interfaces. While new architectures like K2 Horizon promise more efficient performance, the immediate practical frontier is using models as autonomous operators that bridge the gap between legacy software and modern automation.
Key Takeaways
Computer use capabilities are becoming the primary workaround for software that lacks APIs, allowing models to interact with desktop applications like Premiere Pro as a human would.
The K2 Horizon model demonstrates that mixture-of-experts architectures can achieve high-end performance with significantly fewer active parameters, though the industry still lacks a proven, dense baseline for comparison.
Reliability in agentic workflows is currently being solved through manual engineering, such as handoff systems that maintain state across multiple stateless model wakeups.
Direct computer control is not a universal replacement for programmatic integration; CLI and MCP remain superior for speed and reliability when an API is actually available.
A new standard of transparency is emerging in model development, as evidenced by labs choosing to audit and lower their own benchmark scores to provide more accurate performance data.
Claude's 'record a skill' feature can learn workflows, preferences, and some judgment by watching screen recordings and verbal narration, without requiring written prompts. The key insight is that demonstration is a more effective way to transfer tacit knowledge—the kind of expertise that experts display but struggle to codify. This shifts the AI usability advantage from prompt engineers to domain experts who know the work.
K2 Horizon MUA's mixture-of-value-experts architecture routes 4 of 64 value experts per word, yielding 4B active parameters and a terminal benchmark win over a 550B model. The 75 GB model is not a small local model, but the lab's transparent audit—lowering its own headline score—sets a new disclosure standard. The dense 32B control group is unfinished, leaving the architecture's true value unproven.
GPT-6 Astra's computer use capability removes the need for APIs or MCPs by directly controlling applications and websites, as demonstrated in five use cases including video editing in Premiere Pro and phone mirroring. The speaker's own decision chart suggests using CLI or MCP when possible for speed and reliability, indicating computer use is primarily a bridge for tasks without programmatic interfaces. This is a notable step toward general-purpose AI agents, though the demonstrations are promotional and not independently verified.
GPT-6 Astra can be turned into a 24/7 stock trader by setting up scheduled tasks in Codex that connect to Alpaca for trading, with continuity maintained through progress logs and handoff messages between stateless agent wakeups. The presenter demonstrates a 7-day challenge with $10,000 real money, using six daily wakeups at specific market times, and shows how to isolate the trading project to avoid context pollution. The real innovation is the handoff system that makes multiple stateless runs feel like one continuous agent.
Developers are building clever handoff mechanisms that allow stateless models to maintain continuity across long-running tasks, suggesting that we can achieve persistent agent behavior without needing expensive, stateful infrastructure.