Article of the year — vividly demonstrating the new AI-native organizational model beyond OpenAI and Anthropic. Small team + lots of agents, agents-first, with full context & loops. - Spond is a massively multiplayer online game creation platform. The team has only 10 people, but their monthly token spend is $3.2 million, making them one of Anthropic's biggest customers. - From day one, they had a strong sense of context: all recordings and code were archived, waiting for agent capabilities to become strong enough to leverage. - Before fable5 launched, a fleet of always-online agents was simply not possible. The bitter lesson shows up again — when models are strong enough, they crush all engineering patches. GPT-5 was overly cautious, Grok wasn't smart enough, Gemini was poor at long-horizon tasks, and Anthropic's models weren't ambitious enough. - A very aggressive strategy: eliminating agent communication bottlenecks by deleting PRs. The rationale: agents don't specialize like humans do. They have a deep understanding of the entire system, so communication serves no practical purpose. Production stability and development speed are balanced through full test runs every 15 minutes, robust rollback capabilities, and tiered controls (sensitive operations like payments get extra handling). Out of 120,000 changes in the past year, only 0.9% required rollbacks. - They built self-healing software systems. In traditional workflows, agents run on test machines and wait for humans to provide bug reports, creating a massive bottleneck. With the new approach: issues are detected within 8 minutes of affecting players, named and classified within 1 minute, fixed within another 11 minutes, and the platform then runs stably for 4 hours. - Bottleneck shifted — further deconstruction of the CEO's responsibilities. Other parts of the pipeline became so fast that problems were being generated faster than the CEO could make decisions. Waking up to seven hundred questions waiting for an answer. Fortunately, there was a rich history of decision records, so they trained an agent-based judge on historical material to filter out which decisions truly needed human input and which could be answered directly based on precedent. The number of decisions requiring human judgment dropped by an order of magnitude. - Repositioning human decision-making: taste and ambition. Engineering capability is cheap, and demand is booming, but knowing when to say no is still hard. Have enough ambition to imagine every possibility — don't settle for 2% improvements, I want 100x gains. "I don't care how you do it, that's your job, Claude":) https://x.com/felixrieseberg/status/2110023283169247623… — perfectly corroborated by the recent post from Claude's engineering lead. - Why this system is hard for big companies to replicate, and conversely, how to replicate it: 1. You must build loops. The key human job is eliminating bottlenecks in the agent loops. 2. Be willing to take risks. Do you dare to let agents into production while keeping the blast radius under control? 3. Design your org structure around agents rather than directly copying human org structures — that means organizing agents around context, not fantasizing about some ancient three-department-six-ministry system. --- https://x.com/bylixing/status/2113659227122336022… This is an article I wrote a couple of days ago about how to leverage stronger model capabilities in organizational building through context & loops. But since the physical world still has many constraints, what I can practice in reality is essentially a second-tier solution; whereas in the software industry, you can actually go straight for a first-tier solution. It's really gratifying to see such a real-world, frontline example of a first-tier approach.