Rex Ma

Excited to see X-Scientist launch! Proud to be one of the core builders of its harness alongside an incredible team @Xaira_Thera. Here’s what building it taught me about domain-specific agents 👇 1, The model is general. The harness is where domain knowledge lives. Frontier models already “know” a lot of biology. What they lack is the scaffolding to think like a drug hunter: which evidence counts, what order to reason in, when to change course. Our key design choice was to separate scientific thinking from execution. Once coordination and verification are their own layer, they become something you can optimize, and that’s where domain knowledge gets encoded. 2, Tools > prompts. A drug discovery agent needs tools built for its domain. We built ours around our own models like X-Cell and X-Design, alongside bioinformatics and scientific analysis tools. Each tool has to return outputs the model can reason over, not raw dumps. 3, Your verifier is your moat. Coding has unit tests. Biology has noisy data and slow wet-lab loops. Deciding what “good” looks like at each step, and checking it automatically, is what keeps an agent grounded instead of drifting into confident nonsense. It’s also what lets scientists trust its output and step in where it matters. 4, The thinking layer learns. Because scientific reasoning is separated from execution, it can improve on its own. Every run leaves a trace of which strategies produced real evidence and which burned effort. We feed that back into how the agent plans, allocates and verifies. Each run makes the next one a better scientist. General-purpose agents will keep getting better. But in high-stakes science, the last mile belongs to whoever builds the best domain harness. 🚀

This is MCP for harnesses. the model used to be the thing you picked, now it's the harness. With this you can quickly find the most cost-efficient harness for your task and stop paying the harness tax teapaw.org/akshay_pachaar/status/…

Agents built for productivity are converging on one design. From DeepSeek’s Superagent to personal assistants like Grok Bot, the shape is the same: a persistent, proactive agent up front, fanning work out to parallel workers. Our X-Scientist was designed around this from the start. A lead agent handles coordination and verification, and persistent workers do the execution. If the architecture is becoming standard, where does a custom harness get its edge? Sharing my thoughts in two points: 1. The "router" is where domain knowledge belongs. In DeepSeek’s Superagent, the router is deliberately thin. It’s the single front door for every human interaction. It hands each task to an isolated worker and keeps the global context, but it never executes or decomposes work itself. A light model is enough for that. X-Scientist’s lead agent plays the same coordinating role, plus one critical addition: verification. It checks each stage against explicit evaluation criteria. Then it decides whether the evidence supports continuing, trying another approach, or revising the plan. When several workers attempt the same task, it picks the result the evidence best supports. That shows where a domain-specific harness should invest. Execution is becoming a commodity, because any strong model can run a subtask. What doesn’t transfer across domains is judgment: what’s worth doing next, and what counts as “done.” That judgment belongs in the router layer and things can be creative with "router". 2. Human-in-the-loop: make human attention go further, don’t remove the human. Deepseek's blog points out that human attention has become the bottleneck on what agents can deliver. However simply cutting the human out isn't the fix for many domains. In drug discovery specifically, the human decision is the most important thing to keep. Scientists have to make calls on limited evidence, and those calls decide where months of work go. I see two directions: A) Ask at the right moment. Treat human attention as the scarcest resource in the system and schedule it carefully. Interrupt when a human answer is worth the most: high uncertainty, high cost, or a decision that can’t be undone. Batch or defer the low-stakes ones. B) Let expertise compound. Every time a scientist makes a call, the reasoning behind it usually gets lost in a meeting or a Slack thread. A harness can capture those decisions inside the agent loop and use them to train decision models or domain LLMs. Those models can then handle the recurring calls the way the team would. Scientists spend their attention on the decisions that are actually new. The architecture is converging. The edge is in how much domain judgment the harness can hold: in the "router", deciding what’s next and what counts as done, and in the loop with the human, deciding when to ask and learning from every answer. Get both right, and an organization stops renting general intelligence and starts owning its own. Read more on our X-Scientist blog: xaira.com/news/x-scientist-r…

X-Automate + X-Scientist for closing lap in the loop 🧪

Full technical blog: xaira.com/news/x-scientist-r…

More AI investment should go to Canada!