The Unbundling and Rebundling of Agents imo it’s incredibly underrated how important this is and other methods are that build specialized vertical capabilities that plug into harnesses Harnesses are Platforms the goal is to build the best agentic system for your task by plugging in modular components that combine to hit the key design points of accuracy, latency, cost specialized context assembly systems/models are a big part of that: - purpose built agentic search models like SWE-1.5 from @cognition and Context-1 - specialized agentic web search APIs like from @p0 - specialized subagents that do X thing well in a domain we have one notion of an agent that fully bundles all capabilities into a single model + tools to do everything but the future is probably more modular the more exciting future is Open Harness + Modular Specialized components supporting a frontier level model And there’s a an exciting opportunity for companies or open models to win a layer here like for contextual retrieval For vertical tasks, it’s clearly possible to train or harness engineer a model to be great at that task I’m excited to push the frontier of Open Harnesses that let builders plug in and orchestrate specialized layers/models for their tasks like I bet many of the vertical specific frontier agentic systems will be composed of open submodules (Context-1 style) that a frontier model (may also be open like GLM5) will orchestrate. And great companies like @PrimeIntellect will help builders create those specialized sub-modules if they don’t exist through finetuning And we’ll all benefit from lower costs and a bigger opportunity to customize for what we care about and plug that into an open harness exciting times ahead :)