The Cayu blog
Notes from building
and deploying agents.
Perspectives on agent systems, enterprise delivery, and the economics of useful work.
Agent vs Copilot
Copilot, agent, AI-native, and agent-native are not interchangeable. A copilot should be measured by employee productivity. An agent should be measured by the outcome of the job it owns.
ExploreCayu — everything you need to build long-horizon agents
In the physical-world, long-horizon agents are a must-have. Cayu Runtime, Cayu Cloud, and Cayu Compound are the framework to build, deploy, and improve them.
ExploreFDE or not?
In 2026, startup FDEs wear four hats: problem scoper, problem solver, product and agent builder, and momentum generator — backed by an agent-native software factory.
ExploreGateway or Not?
AI cost should be measured in dollars per outcome, not just dollars per million tokens. Gateways that tag each model call are the wrong unit of control for long-horizon agents.
ExploreTo Agent or Not?
If you can specify the procedure, build software. If the system needs judgment to determine what the procedure should be, use an agent.
ExploreToken cost: $/1M tokens? No!
The lazy answer for measuring AI cost is dollars per million tokens. Different AI workflows need different KPIs — and the metric you use to control spend is not the same one Finance uses after the work finishes.
ExploreWait. Remember. Recover.
Access to models is not the bottleneck. The actual problem is putting agents into production to complete jobs end-to-end, create measurable value, and compound knowledge.
Explore