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.
Choosing an approach
Section titled “Choosing an approach”| Situation | Use | Why |
|---|---|---|
| One task, one train of thought | One conversation | Shared context is the whole value; splitting it costs you continuity |
| Unrelated tasks — a bug fix and a doc update | Separate conversations | No shared context needed, and each keeps its own history and folder |
| Comparing how two agents tackle the same prompt | Separate conversations, same folder | You read both answers side by side and keep the better one |
| One goal, several roles — implement, review, test | A team | A lead splits the work, assigns it, and synthesises the results |
| Many similar items — a list of files to migrate | A team | Members take items in parallel under one coordinator |
What runs at once
Section titled “What runs at once”- 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.
Why some work is deliberately sequenced
Section titled “Why some work is deliberately sequenced”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.
Keeping cost sane
Section titled “Keeping cost sane”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.