PRACTICE · SYNVOX
An AI agent team: roles, a shared chat, and project memory
Several models should not all answer every user message. A team helps when members have distinct responsibilities and a lead agent assembles one result in a shared chat. This alpha SYNVOX workflow is still evolving and should be validated on your own tasks.
A lead agent and helpers have different jobs
The lead agent receives the user task, decides whether help is needed, and owns the overall outcome. An analyst studies requirements and constraints, a developer makes changes, and a reviewer finds issues in the result. Simple questions need no helpers. Useful delegation has a concrete subject and expected output, rather than asking everyone for an opinion.
Pass memory without replaying the whole chat
A helper normally needs the goal, relevant files, decided constraints, and verification criteria. A complete transcript can add noise and stale assumptions. Long-term memory should contain a short confirmed decision plus a reference to a file or check. Draft reasoning and unverified conclusions should not become project facts.
When several agents get in the way
Two developers changing the same file can create conflicts. A reviewer without the changes can review the wrong subject. Divide ownership by files or tasks and do dependent steps in sequence. Every extra agent adds a request, waiting time, and provider load; one agent is often more convenient for a small change.
An example task for AI
Lead agent, organize a fix for search in this project. Give the analyst the requirements and bad-result examples; give the developer the agreed conclusion and relevant files; give the reviewer only completed changes and test results. Do not start helpers without a separate task. Assemble the final result here with authors and remaining risks.
This is a task-framing example, not a report of completed work.
Common questions
Does each helper need a separate user chat?
In the SYNVOX team workflow, participants work in the current chat with assigned roles. That differs from separate independent chats: delegated context should remain connected to the common goal.
Are multiple agents and multiple accounts the same thing?
No. An agent is a task participant with a role. An account is a provider access session. Multiple accounts still require distinct tasks, isolated context, and compliance with service limits.