Agents & Subagents
The main agent can delegate work to subagents that run in the background while you keep chatting. Long research, multi-step builds, parallel investigations — the agent fans out and reports back.
How it works
- Ask for something long-running — the agent decides when a subagent helps (or you can ask explicitly).
- Background execution — subagents work while the conversation continues; you get the result when it's ready.
- Shared memory, single-writer — subagents share the conversation's memory safely; only one writer at a time, so nothing corrupts.
- Per-call model & tools — each subagent call can pick its own model and tool set. Cheap fast model for research, strong model for the final answer, browser tools only where needed. This keeps costs and context under control.
Why subagents matter
Without them, one long task blocks the whole chat and burns the main model's context on every step. With them:
- Research runs in parallel (three investigations at once, not one after another).
- The main conversation stays responsive and its context stays clean.
- Each subtask gets the right-sized model instead of the most expensive one for everything.
Agent modes & workspace scoping
- Plan / Build modes can be scoped to a project folder: when you pick the mode, the app asks which folder inside the workspace it may touch.
- The agent scans the whole workspace only when you explicitly ask — scoped modes can't wander into unrelated projects.
Context hygiene
- Automatic Jev checkpoints compact only model-visible context; the full chat history is always retained in the app.
- If context grows fast during normal chat, check Settings → Diagnostics — the log shows exactly what grew it, so the cause is traceable instead of mysterious.