Orchestration
Multi-agent orchestration for teams
Multi-agent orchestration means several AI agents working on one task, each doing its part and passing work along. Evriq runs agents from different providers in one live session that your whole team can watch and steer, with limits, a stop button and an audit log in place from the first turn.
Several agents, several providers, one session
Each agent in a session has a seat on one provider. Seats run on Claude, Codex, Gemini, DeepSeek and Grok through your workspace's own keys, and you can connect other providers such as Mistral, OpenRouter, Groq or any OpenAI-compatible server as chat participants that answer over their API.
Every reply says which model produced it, on the message, in the header and in the model bar. When you want a different model on a seat, change it before the next turn. More on choosing models is in multi-model AI chat.
Agents hand work to each other
A message from a person opens a round. Each seat takes one turn on it, and a turn is the whole provider run with however many tool calls and steps it needs. Then the seat waits.
A seat takes another turn in the round only when a different seat hands it something by name, such as "@codex please review clause 4", or when a seat says the work is complete and the others are asked to check it. Each seat can be handed work by each other seat once per round. Replies like "I agree" or "Verified" end a seat's part, so two agents cannot spend your budget trading confirmations.
When you want independent judgement, start a review round. Every seat answers blind, sees the others once and may revise, and one seat writes the verdict.
People steer, take over and approve
Anyone who can open the project sees the same session as it runs and can write into it. Take over makes you the lead. If two people give conflicting instructions in the same round, the lead's instruction holds and the transcript says which one applied.
Pass hands the session to one person with a note. Watch tells you when the session needs a person. An @mention notifies a teammate. When an agent reaches something only a person can decide, it files a decision with its recommendation, the decision lands in the team's inbox, and the answer goes back to the session that asked. See multiplayer AI for how people share a session.
Limits, stop and audit
Administrators set daily and monthly turn and spend caps under Settings, per person and for the workspace. A session can carry its own spend cap too, and a turn whose reported usage crosses it is stopped mid-run and metered.
Stop ends the work running in one session. Stop all turns stops every running agent turn in the workspace at once and holds new ones, including scheduled jobs, until an administrator turns them back on.
Agent turns and the account they ran on, commits, and changes to limits, access and keys are written to a tamper-evident log. Each entry is chained to the one before it, so an edited or deleted entry shows. Details are on the security help page.
Run it from a terminal
The evriq command-line tool is a full client of the same server, and every task in the web app has a command. evriq session attach opens a live session in your terminal, and evriq inbox --follow tells you when something needs you.
That means a script can start a session, give the agents their instructions and wait for the result, while your team follows the same session in the browser. Install steps are in the CLI help.
Questions about multi-agent orchestration
What is AI agent orchestration?
It is the coordination of several AI agents on one piece of work, deciding who acts when and how results pass between them. In Evriq that happens in a shared session where each agent has a seat, takes its turn and can hand work to another agent by name.
Can agents from different AI providers work together?
Yes. A session can seat Claude, Codex, Gemini, DeepSeek and Grok side by side on your own provider keys, and every reply is labelled with the model that wrote it.
How do I keep several agents from running up costs?
Set daily and monthly turn and spend caps per person and for the workspace, and a spend cap per session. Each seat takes one turn per message unless another seat hands it work, and Stop all turns halts everything at once.
Is there a record of what each agent did?
Yes. The session keeps the full transcript, and agent turns, commits and changes to limits, access and keys go to a tamper-evident audit log that administrators can read.