Harnesses often get dismissed as just scaffolding, just prompt engineering, and not real research. But that couldn't be farther from the truth. The same model weights that score 30% on ARC-AGI score 95% with a better harness. So we gathered a group of researchers and founders working at the frontier to do a deep dive into the state of harnesses. We cover how we got to this point, the case for making your harness as expressive as possible, and what YC learned building an agent for every employee in the company. 00:00 - @FrancoisChauba1: Why harnesses matter 04:27 - Building an auto-researcher by accident 07:13 - A five minute history of harnesses 13:56 - Self-improving harnesses 18:35 - @sethkarten: Prime Agent, a self-improving RLM harness 21:50 - Context as an L1, L2, L3 cache 24:51 - From Turing machine to von Neumann computer 28:33 - Messaging between agents 30:04 - ARC-AGI results 33:09 - Emulator Bench and GPU kernels 37:30 - @JonSaadFalcon: OpenJarvis, personal AI on personal devices 38:26 - How far behind are local models 39:21 - The five primitives of a personal AI stack 42:47 - Letting cloud models optimize your local stack 43:53 - 800x cheaper than the cloud 45:58 - @josh__france and @jbellregan: QM, YC's agent harness for work 47:29 - A history of YC's internal agents 49:24 - OpenClaw and a fleet of 50 agents 51:04 - Pulling the brain out of the sandbox 54:43 - Letting the agent choose its own sandbox and model 57:16 - The grind tool: budgets on goals 58:50 - Agents don't understand social context

Some new misalignment disclosures from OpenAI: • Last Sunday morning, one of our models was able to gain unauthorized access to the internet during RL training (~all inference for our most capable models remains stopped until we have hardened our systems further) • In May, a version of HPIM uploaded a employee's GitHub token to the internet, causing the model to be quarantined for two weeks • A new research finding, demonstrating that one can construct self-replicating prompt injections alignment.openai.com/misalig…

🚨 What if your LLM team could dynamically reorganize itself mid-task? Introducing LATTE: a framework that gives teams a ~caffeine~ boost via an evolving task graph they build and adapt in real time, achieving SOTA performance with fewer tokens and conflicts. New paper 🧵👇

Some data we recently assembled on entrepreneurship/compute in Europe: eudata.vercel.app. We hope that one of the useful roles that Stripe can play is in collecting and publishing empirical data pertaining to entrepreneurship and industry in Europe. There's growing appetite to get Europe on a better footing, and cross-sectional comparisons can often shine light on where opportunities lie. If you're interested in this kind of thing, we publish more at stripeeconomics.substack.com.

prime agent v0.9.6 is out: ◆ Support for GPT-6 Sol, Opus 5.5, and Grok 4.7 ◆ /mcp plugin catalog with one-click connections to Linear, Notion, Posthog, Stripe, and 60+ more services ◆ Huge perf and reliability pass 🫡 Lots more coming soon :)

Astra has randomly gained the ability to drive a car, seemingly without specialist training.

Introducing Prime Sandboxes: MicroVM sandboxes purpose-built for RL training. Model training requires running tens of thousands of concurrent sandboxes, leading to complex and costly configuration. We built Prime Sandboxes for our own team. Today we're releasing them publicly.

Using open benchmarks has gotten very frustrating because it's very hard to disentangle what the model already knows / is actually bottlenecked by. E.g. when testing RLMs on "new" long benchmarks, some models will just magically regex for the right offloaded information and grab the solution... This is roughly how I've interpreted the HarnessTax paper as well, where performance seems to basically be a function of only the model capability and not the harness, but in practice we observe very different behavior / preferences with different harnesses. Model capability dominates on open benchmarks specifically. Not really sure what this implies because newer benchmarks constantly get swallowed, but I hope someone comes up with a better way to benchmark models in the open because it's largely uninteresting atm. And no, the solution isn't a benchmark that tests how far models get in a 2-day task. There's definitely a better way to host and/or manage benchmarks to actually gauge capabilities, at least for longer than a year.