Q
Subagent or one focused session?
Delegate when the work is bounded, parallel and read-heavy, research, audits, blast-radius checks. Keep the edit loop inline when full context pays.
Docs · Subagents & Parallel Work
Subagents & Parallel Work.
Fresh windows for dirty work.
Direct answer
Subagents run focused tasks in their own context windows and return summaries, so exploration sprawl never enters the parent's working memory. Anvaya ships four profiles with per-type round and time budgets, up to four concurrently, stale-heartbeat detection, and live panes you can drill into from the TUI.
Subagents & Parallel Work
The four profiles
01explore · read-onlyFast reconnaissance: map a subsystem, find callers, gather evidence. Read-only tools, short budgets.
02plan · read-onlyProduce a step plan and risk notes for a task, useful before approving a mutating run.
03build · full toolsBounded implementation work in its own window, with write and exec access subject to the sandbox and the parent's approval policy.
04general · full toolsThe default for mixed tasks that need both investigation and edits.
Subagents & Parallel Work
Operating rules
01Four at a time, then queueConcurrency is capped at four; additional delegations queue rather than fail. Budgets per type keep one delegation from eating the session.
02Heartbeats, not hopeEvery running subagent carries a heartbeat; thirty seconds of silence marks it STALLED in the pane so you can inspect, cancel or reissue.
03Return contractsAsk for the shape you need back, a list, a verdict, a patch summary. The parent receives the summary, not the transcript.
04No recursion by designSubagents cannot spawn subagents or ask the user questions; those tools are removed from their set, which keeps delegation bounded.
Questions
Asked About Subagents & Parallel Work.
Q
How much does delegation save?
Exploration tokens live and die in the child window; only the summary enters the parent. Savings scale with how much reading the task requires, and the per-type budgets cap the downside.
Q
Can subagents edit files?
The explore and plan profiles cannot; build and general can, subject to the same sandbox policy and approval gates as the main agent.
Run Agents That Fit On Your Laptop.
25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.
Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.