Every era of computing has mistaken effort for value. First we paid for lines of code, then for hours billed, and now for tokens. That's because every system eventually optimizes for what it can measure. When tokens became the target for AI, so we got agents that consume heavily and deliver unevenly. Routing decisions through @levantolabs' Sage fixes this. Every task is weighed before any work starts and routed to the model it deserves. This way no energy is wasted

lfggggg amazinggg

OpenAI’s Agents API is insane. We’ve been working privately with the @OpenAI team for the last couple of months - researching it on real workloads, debugging edge cases, and building on top of it. One idea we've been thinking for a while: Agents should scale dynamically with the workload of the org. To test this idea, we made an open-source project that is completely managed by AI agents. In the demo, quirqy gets a feature request in chat, understands what needs to happen, creates the feature request, assigns an agent, spins up a fresh sandbox, builds + tests the change, opens the PR, and then shuts everything down when the job is done. Now imagine that across a company. 10 tasks come in → spin up 10 agents. Work disappears → scale them back down. No sandboxes burning money doing nothing. The agents scale with the work, and the execution environments scale with the agents. For the workloads we’re testing, this can cut the infrastructure cost of running agent fleets by up to 50% compared to keeping runtimes alive all the time. That’s the big idea behind Orgs on XO: a managed layer where agents, tools, and environments show up when the work shows up. Really fun building this with the @OpenAIDevs team and really grateful for the support. It feels like one of the missing primitives for operating large fleets of agents without having to build all of the infrastructure yourself. The Agents API feels like a huge step toward making large-scale agent systems actually practical to run.

Agents API is super fun to build on top of, It feels like one of the missing primitives for operating large fleets of agents without having to build all of the infrastructure yourself. Saved us 1000s of dev hours, usually spent around provisioning, security and configurations, before the agent is ready for production. Thanks for the early access @OpenAIDevs @_apoorvjha

OpenAI’s Agents API is insane. We’ve been working privately with the @OpenAI team for the last couple of months - researching it on real workloads, debugging edge cases, and building on top of it. One idea we've been thinking for a while: Agents should scale dynamically with the workload of the org. To test this idea, we made an open-source project that is completely managed by AI agents. In the demo, quirqy gets a feature request in chat, understands what needs to happen, creates the feature request, assigns an agent, spins up a fresh sandbox, builds + tests the change, opens the PR, and then shuts everything down when the job is done. Now imagine that across a company. 10 tasks come in → spin up 10 agents. Work disappears → scale them back down. No sandboxes burning money doing nothing. The agents scale with the work, and the execution environments scale with the agents. For the workloads we’re testing, this can cut the infrastructure cost of running agent fleets by up to 50% compared to keeping runtimes alive all the time. That’s the big idea behind Orgs on XO: a managed layer where agents, tools, and environments show up when the work shows up. Really fun building this with the @OpenAIDevs team and really grateful for the support. It feels like one of the missing primitives for operating large fleets of agents without having to build all of the infrastructure yourself. The Agents API feels like a huge step toward making large-scale agent systems actually practical to run.

🎉🎉🥂 always grinding @sydneylai 🫡

Thank you team for having me over

Most businesses today can see how much AI costs, but still struggle to understand what it actually delivered. Token bills tell us what the machine consumed, not whether the work created value. That gap is what we're trying to address with quirq, a unit I’m working on to measure verified business impact from Agentic work. The future I see is not one where AI replaces human accountability or talent, but one where humans stay accountable for defining the outcome, deciding what it is worth, and setting the budget. AI then goes and executes the work, while the environment verifies what was actually done, what it cost, and where humans still had to step in. This becomes especially important as companies move from AI demos to actual agentic workforces. My belief is that the next phase of AI adoption will not just be about better models, but better environments that can prove what work was done, what it was worth, and whether the system is getting better over time. We wrote a paper addressing this problem and proposing a new way to measure the business impact of agentic work. Sharing an early draft and would love feedback, criticism, and sharp counterarguments.

Most businesses today can see how much AI costs, but still struggle to understand what it actually delivered. Token bills tell us what the machine consumed, not whether the work created value. That gap is what we're trying to address with quirq, a unit I’m working on to measure verified business impact from Agentic work. The future I see is not one where AI replaces human accountability or talent, but one where humans stay accountable for defining the outcome, deciding what it is worth, and setting the budget. AI then goes and executes the work, while the environment verifies what was actually done, what it cost, and where humans still had to step in. This becomes especially important as companies move from AI demos to actual agentic workforces. My belief is that the next phase of AI adoption will not just be about better models, but better environments that can prove what work was done, what it was worth, and whether the system is getting better over time. We wrote a paper addressing this problem and proposing a new way to measure the business impact of agentic work. Sharing an early draft and would love feedback, criticism, and sharp counterarguments. xo.builders/whitepaper