Sub-Agents: A Research Team on One Phone

A sub-agent is a second assistant with a narrow job, its own prompt and its own small set of tools. The main assistant hands it a question and gets a structured answer back. Several can be working at once.

The five that ship

• Researcher — answers one sub-question, searches, reads at most two pages, returns facts with sources.
• Fact Checker — takes a claim and returns supported, refuted or unclear, with evidence.
• Summarizer — condenses text. No tools; pure reasoning.
• Document Analyst — searches your own documents and cites them.
• Extractor — pulls structured data out of unstructured text, against a schema you supply.

How a phone runs a team

This is the interesting constraint. A phone has one model context, and loading several would end the app immediately. So sub-agents parallelise the part that is waiting — searching and fetching pages overlap freely — while every generation step queues in strict order on a single lock shared with the main chat. You get the wall-clock benefit of concurrency without ever needing a second model in memory.

Bounded on purpose

How many agents may run at once comes from BatteryGuard: four while charging, fewer on battery, none when the phone is low or hot. Each run has a hard token budget, each agent a ninety-second timeout, and there is a Stop button throughout. An agent framework without those is a good way to discover your phone at 4% an hour later.

Where it goes next

The seed agents are editable and you can write your own. For a full investigation with a written, cited report rather than a delegated sub-question, see deep research.