Glossary · Subagents & Isolation

Subagents & Isolation.

Fresh windows for dirty work.

Direct answer

A subagent handles a focused task in its own clean context window and returns a 1–2K summary; the parent's working memory never sees the exploration sprawl. This Isolate strategy prevents context multiplication (full-history hand-offs that explode tokens and confuse children), enables parallel investigation, and is the standard treatment for research spikes, blast-radius analysis, and multi-file reconnaissance.

In Anvaya

How We Implement It.

01Subagents & IsolationAnvaya ships 4 subagent profiles with green panes, heartbeats, STALLED detection, and drill-down — plus swarm packs for multi-agent orchestration with merge gates.

Questions

Asked About Subagents & Isolation.

Q

What should I delegate vs do inline?

Delegate bounded research with clear return contracts (lists, summaries, verdicts). Keep tight edit loops inline where full context pays.

Q

How do subagents save tokens?

Tens of thousands of exploration tokens live and die in the child window; only the distilled summary enters the parent. Per delegation, often 10–20x.

Q

What is context multiplication?

Passing full parent history into children (and grandchildren) — token explosion plus confusion from irrelevant history. Pass goals, not transcripts.

Stop Starting From Zero.

One binary. 11+9 Rust crates. 545 tests. Hand-written HNSW index. Three transport modes. Four providers, Ollama, Anthropic, OpenAI, Siemens. Zero API keys required to start. Mind remembers everything after the first session.