“Build now, but build for production loops, not demos. We see the biggest returns from agents that run long enough to learn from outcomes, not from waiting for a better base model.”

“Build now, but build around the failures. We started Instinct because the gap in τ-bench was obvious: even GPT-4o cleared fewer than half of real-world tasks, and pass^8 dropped below 25% in retail. The teams that win will be the ones that learn the shape of those failures now, not the ones that sit out the year.”

“I would start now, but start small. I got back into coding over the holidays by picking low-stakes, for-fun projects first, and by the time I was building HyperAgent itself I had the confidence that agents could actually do the work. The people who wait for agents to just work are usually the ones who never learned what to ask them to do.”

“Build now, but build where agents are already working rather than where you need a leap. At Cresta we went live with small deep nets in 2018, years before LLMs, because call centers had a constrained supply of human agents and a clear task to automate.”

“I’m building agentic teams now, but I keep the architecture simple: a main agent with subagents, minimal tools, and let them manage their own memory rather than bolting on every new framework. The current stack is noisy and expensive, yet the teams that figure out orchestration today will be the ones that don’t have to relearn it next year.”

“Build now, but pick a narrow use case where the models are already good enough. A year ago Ava barely worked; the past six months flipped customer outcomes for us because the weak links in outbound, not the whole stack, got better. Waiting a year means someone else learns the messy integration work you're avoiding.”

“Build now, but choose the right surface. When I started building agents in 2023 and 2024, tool calling, memory, and orchestration were genuinely painful; today the frontier models and frameworks have moved enough that the bottleneck is usually your evaluation pipeline, not the agent itself.”

“I spent the last 18 months running four machines of AI subscriptions to ship Vois and build WorkflowOS, and the real cost wasn't the tools, it was learning to manage agent coordination before the models got better. Start now, but build around state and intent, not the current model, because execution is getting cheap and your attention is the actual bottleneck.”

“I built ClawdBody in the middle of the agent hype cycle and got 5,300 signups in 38 days because the tooling was just good enough to ship something people wanted. The thing that matters is not whether agents work in a year, it is whether you can build a workflow that works today and keep it alive as the models improve.”

“Build now, but build for production loops, not demos. We see the biggest returns from agents that run long enough to learn from outcomes, not from waiting for a better base model.”

“Build now, but build around the failures. We started Instinct because the gap in τ-bench was obvious: even GPT-4o cleared fewer than half of real-world tasks, and pass^8 dropped below 25% in retail. The teams that win will be the ones that learn the shape of those failures now, not the ones that sit out the year.”

“I would start now, but start small. I got back into coding over the holidays by picking low-stakes, for-fun projects first, and by the time I was building HyperAgent itself I had the confidence that agents could actually do the work. The people who wait for agents to just work are usually the ones who never learned what to ask them to do.”

“Build now, but build where agents are already working rather than where you need a leap. At Cresta we went live with small deep nets in 2018, years before LLMs, because call centers had a constrained supply of human agents and a clear task to automate.”

“I’m building agentic teams now, but I keep the architecture simple: a main agent with subagents, minimal tools, and let them manage their own memory rather than bolting on every new framework. The current stack is noisy and expensive, yet the teams that figure out orchestration today will be the ones that don’t have to relearn it next year.”

“Build now, but pick a narrow use case where the models are already good enough. A year ago Ava barely worked; the past six months flipped customer outcomes for us because the weak links in outbound, not the whole stack, got better. Waiting a year means someone else learns the messy integration work you're avoiding.”

“Build now, but choose the right surface. When I started building agents in 2023 and 2024, tool calling, memory, and orchestration were genuinely painful; today the frontier models and frameworks have moved enough that the bottleneck is usually your evaluation pipeline, not the agent itself.”

“I spent the last 18 months running four machines of AI subscriptions to ship Vois and build WorkflowOS, and the real cost wasn't the tools, it was learning to manage agent coordination before the models got better. Start now, but build around state and intent, not the current model, because execution is getting cheap and your attention is the actual bottleneck.”

“I built ClawdBody in the middle of the agent hype cycle and got 5,300 signups in 38 days because the tooling was just good enough to ship something people wanted. The thing that matters is not whether agents work in a year, it is whether you can build a workflow that works today and keep it alive as the models improve.”