Skip to content

Working with several agents

Pavii.Ai runs one agent process per conversation, and conversations are independent. Nothing serialises them, so several agents genuinely work at the same time.

SituationUseWhy
One task, one train of thoughtOne conversationShared context is the whole value; splitting it costs you continuity
Unrelated tasks — a bug fix and a doc updateSeparate conversationsNo shared context needed, and each keeps its own history and folder
Comparing how two agents tackle the same promptSeparate conversations, same folderYou read both answers side by side and keep the better one
One goal, several roles — implement, review, testA teamA lead splits the work, assigns it, and synthesises the results
Many similar items — a list of files to migrateA teamMembers take items in parallel under one coordinator
  • Each conversation holds its own agent process. There is no global cap on how many run together — the practical limits are your machine and your provider’s rate limits.
  • Within a team, each member has its own queue and runs at most one batch at a time, but members are independent of each other.
  • Work waiting on a busy member is queued rather than dropped, with user messages taking priority over automated coordination traffic.

When one task depends on another, a team lead is instructed to dispatch the first, wait for it to report back, and only then dispatch the second.

This is not a scheduler limitation — it is about how model providers work. An agent that sits “waiting” inside an open request eventually hits the provider’s request timeout and is marked failed. Finishing a turn and being woken again later is both cheaper and more reliable than holding a connection open.

The practical upshot: independent work fans out, dependent work goes in order, and you should describe dependencies to the lead rather than assuming it will infer them.

Every parallel agent is a separate stream of tokens. Before fanning out, it is worth checking the context meter in the composer and the usage figures on your profile — see Tokens and usage.