TinyFish brings research planning into its main reasoning loop

TinyFish's July 18 engineering account describes a research agent that gathers web evidence and writes cited reports. A separate planner received a compressed state instead of the orchestrator's accumulated conversation. When users narrowed their question, it could broaden the research again. The team moved planning into the orchestrator, the component directing the investigation, so both used the same context. This is an account of that system, not a universal comparison of agent designs.

The company's architecture guide recommends defining required facts, acceptable sources and freshness before searching. It calls this an evidence contract, explicitly a working term rather than an industry standard. Search finds unknown sources; direct retrieval fits a known, readable URL. The application then checks whether evidence covers the question, remains incomplete or conflicts. Reaching a time or cost boundary should produce an explicit limitation or partial result. For developers, the proposed control is therefore about deciding what counts as completion, rather than treating any fluent answer as sufficient. Unresolved questions should remain visible.

MarkTechPost's October 23 article brings these guides together and discloses TinyFish sponsorship. Its discussion draws on the company's materials; it should not be treated as independent confirmation of the system's performance.