Skip to main content
  1. Check whether you need a custom Expert. Ask Cosmos Advisor for a Template Expert from the Cosmos knowledgebase before building your own. Augment maintains the templates and optimizes their cost and quality.
  2. Lean on Cosmos Advisor. Use it throughout setup and iteration to configure Experts, Environments, integrations, and Automations correctly, and get best-practice recommendations.
  3. Prefer one agent per end-to-end task. Cosmos agents can reliably execute long sequences of instructions; splitting one workflow across an orchestrator and multiple workers is usually an anti-pattern that adds cost, coordination bugs, and debugging complexity. Delegate only for a clear need, such as keeping an event-listening agent responsive while independent tasks run.
  4. Optimize quality before cost. Start with the model Cosmos Advisor recommends, typically a mid-range model. Get reliable results first, then test lower-cost options against the same quality bar.
  5. Roll out small and iterate. Test on real tasks, deploy to a few users, and collect feedback. Inspect failures and improve context, prompts, or configuration until the output meets the standard you’d expect from a human doing the work.
  6. Keep humans in the loop. Humans remain responsible for outcomes and judgment calls; agents often lack architectural and business context. Ask for input on choices such as frontend versus backend implementation, or whether backward compatibility matters for a product with no external users.