An agent is a loop with a tool list and a memory. Most agent failures are loop failures or tool-list failures, not model failures.

Planning depth is the silent cost driver. An agent that plans three steps ahead costs four times more than one that plans one step ahead. Plan no further than you must.

Multi-agent systems amplify both the strengths and the failures of single agents. Add a second agent only when the first one has measurable gaps the new one closes.

Tool calls are the agent's hands. A tool that returns ambiguous text is a hand the agent fumbles with. Tools that return structured data are tools the agent uses well.

Memory is the part of the agent the team forgets to design. Without an explicit memory contract, the agent invents one and it is usually wrong.