We doomscroll, you upskill.
Finding signal on X is harder than ever. We curate high-value insights on AI, Startups, and Product so you can focus on what matters.
Page 1 • Showing 1 tweet
Model providers are cracking down on usage in external harnesses. Recently for coding agents. And they have every right to do so. They do not heavily subsidize usage out of the kindness of their hearts. They do so to bring users into their products instead of just using their APIs. Getting your harnesses from your model-dealer makes yourself vulnerable to being one policy change away from a broken workflow. Single provider = single point of failure. This is why you should buy models from model companies and agents from agent companies. Agents from model companies tend to be too rigid. Model companies want you on their agent. So they subsidize compute. Similar to Uber: burn cash to capture the market, figure out unit economics later. The problem with this strategy is that the harness can be built better by others. Agent labs are already doing it. Model companies have no fundamental advantage in building harnesses. Post-training on tool use is not exclusively advantageous. They do have two major limitations: split focus and being stuck with one model family. An agent lab puts everything into the user experience and the agent itself. Supporting multiple models has a compounding benefit. Polishing a rough edge for Model A turns out to fix an issue with Model Y, which Model Y's team never found or prioritized. More time is spent on general techniques (such as compaction, prompt caching, and others) that work across all models. This leaves you with a more well-rounded agent. This is the thesis behind @FactoryAI and what our whole team is 100% focused on. This is how we've maintained our spot as the #1 coding agent, even on vibe-based evaluations. Our plans are all model-agnostic, with no hourly-limits, no "fair use" ambiguity, and BYOK-friendly. If you got hit by recent restrictions, keep enjoying your models, but consider marrying the harness... And as a promo for those affected, let's play a fun game: List price for our Max Plan is $200. We have a discount price, but it's a secret that you must guess. If your guess is ≥ the discount price, you get the full Max Plan at that guess price. If your guess is too low, you're stuck at list price. Closest guess without going over wins. Submit your move here.