Large code migrations, split across parallel agents
Hand Hoplite the whole migration. The main agent plans it, delegates the pieces to subagents that work at the same time, reads every result and reports back, all in one thread you can follow and steer.
What is a multi-agent code migration?
A code migration is one change carried across many parts of a codebase: a framework upgrade, a new router, a replaced library, a renamed module. A multi-agent migration gives that job to an orchestrator, which splits it into independent tasks and hands each to its own agent instead of working through the pieces one by one.
In Hoplite, the agent in your thread is the orchestrator and the agents it delegates to are subagents. You follow them in the thread's Subagents view, message or stop any one of them, and hear from the main agent when the work comes back together.
One plan, many subagents at a time
The main agent breaks the migration into tasks and delegates each one to a subagent. Paid plans run up to eight subagents at once for each parent task, and Free runs four. Extra tasks wait in a queue, show their position and start on their own as others finish.
Every subagent in one view
The Subagents view groups running, queued, failed and finished work, with filters and a layout that works on a phone. Open any subagent in its own linkable tab to follow its tool activity and replies, and read its summary when it's done.
Steer one without stopping the rest
Message a running subagent to steer it after its current step, or queue the message for the end of its turn. The main agent gets a brief note that you did. You can also stop one subagent, or message the main agent while its subagents keep working.
Checked before it's reported
Each subagent returns a report to the main agent. A failed one says why it stopped and hands back any partial summary, so the main agent can decide how to continue. For UI work, the main agent is responsible for checking the combined result in the running app before it claims completion.
One agent in a terminal vs. Hoplite subagents
| Question | One agent in a terminal | Hoplite main agent and subagents |
|---|---|---|
| How the work runs | One piece after another in one session | Delegated to subagents that run at the same time |
| How many at once | One | Up to eight per parent task on paid plans, four on Free, the rest queued |
| Following progress | Scrollback in one terminal | Running, queued, failed and finished, grouped in one view |
| Changing course | Interrupt the whole session | Message one subagent, or the main agent, while the rest keep going |
| When a piece fails | Find it in the scrollback and start again | A failure reason and any partial summary, passed to the main agent |
| The record | Whatever the terminal kept | A transcript export with each subagent's conversation nested |
Migrations with subagents, answered
- How many subagents can run at once?
- Paid plans run up to eight subagents at once for each parent task, and Free runs four. Additional tasks queue instead of being rejected, show their queue position and start automatically when another finishes. Cancelling the parent also cancels its queued tasks.
- Can I talk to a subagent directly?
- Yes. Open it from the Subagents view and send a message to steer it after its current step, or queue one for the end of its turn. Images, PDFs and videos work when its model can read them. The main agent gets a brief note saying you messaged that subagent.
- What happens when a subagent fails?
- Its row shows the reason it stopped, and the main agent receives recovery guidance and any partial summary. Where recovery controls are available, you can inspect the failed attempt and resume it with an optional note. Subagents also save progress as they go, so interrupted work continues from the last saved step.
- How does a subagent's work get back to the main agent?
- A subagent either shares the thread's workspace or works in a separate one. A separate-workspace subagent commits its changes, and Hoplite transfers those commits back to the parent. If the transfer fails, Hoplite keeps the subagent's workspace for recovery instead of deleting the work.
- Can subagents use my MCP integrations?
- No. MCP integrations stay with the main agent, so it does that work itself or gives a subagent material it has already read to analyse.
- Can I message the main agent while subagents are working?
- Yes. A new message ends the main agent's wait so it can answer you or change course, and its subagents keep running with their work intact.
Hand over the whole migration
Connect a GitHub repository, describe the migration and ask Hoplite to split it across subagents. Start on Free.