Technique
What is Planning?
Also known as: task decomposition
Planning is the step where an AI agent turns a goal into an ordered set of sub-tasks before acting, so it can sequence tool calls deliberately rather than improvising one step at a time.
Ask an agent to onboard the new hire and there's no single tool for that. Planning is the agent decomposing it: create the account, add them to the right groups, book the intro calls, send the welcome note. Plans can be made once upfront or revised as the agent learns — plan-and-execute styles do the former, ReAct-style agents do the latter implicitly. The trade-off is familiar: an upfront plan is cheaper and auditable but brittle; continuous replanning adapts but costs more steps.
A visible plan is also a safety feature. In Velaris, a multi-step agent's plan is what tells you a run is about to touch a destructive-tier tool before it gets there, and it's what an approval prompt is describing when it asks you to confirm. Planning pairs with the capability mesh: the agent doesn't need every connector's schema in its prompt to sketch a sequence — it plans in terms of intent, then searches the mesh for the tool that satisfies each step at execution time.