Sponge

Deep research should return a trail of documents

Sponge can trust a deep-research run only when the pages that run used remain inspectable. Tokens may move through cheap workers. The trail of documents still has to survive as documents.

On 22 April 2026, Kavin Anand and Toby Liang published Let the tokens flow. Sail powers efficient, reliable deep research. Sail reports topping BrowseComp-Plus at 90.72% accuracy and $0.15 per query by treating research as compute. A GLM-5.1 orchestrator never reads documents. Cheaper gpt-oss-120b readers ingest truncated hits in parallel, drop noise, and return compacted evidence. hraness recorded that post as a reading digest. This page is not that digest. This page states Sponge’s claim.

A search subagent should return documents, not the answer already argued, from Chroma Context-1, that a search step should return pages. This page starts from a different source. Sail’s BrowseComp-style stack spends tokens so a long trajectory stays reliable. A research product still has to keep the trail those tokens touched.

Tokens can flow without erasing the trail

Sail’s architecture is a cost argument. The orchestrator proposes searches and submits answers. Swarm readers protect the main context so the stack can raise how many documents each search returns. Across 830 queries, most of the 6.5 billion tokens land on those workers.

That persistence is useful. Compacted evidence is not the record. A later person inspects a page, not a swarm summary that replaced it. How Sponge knowledge works keeps citations bound to graph records. The about page calls the durable object the path: identities that survive renaming, statements that retain context, assertions that remain attributable, and editions that preserve their own history. Those objects need the documents the run found.

A one-shot answer is the wrong product object

Sail’s benchmark goal is a single authoritative answer. Sponge’s public contract is different. Agents may search and propose. A person still accepts research into a private graph. A public /k locator still requires a second review, eligible rights, an exact supported edition, explicit publication enablement, and a deliberate release. If the run already wrote the answer, those later acts have nothing inspectable to review.

Retrieval is the product already treats a generated answer as authorship. Token flow can keep a long run alive. It cannot replace the trail a reviewer later walks.

Keep the documents while the swarm spends tokens

Sail is right that bulk parallel ingest, not a custom retriever, is the compute bottleneck. A research workspace can let tokens flow. It still has to retain the pages those workers read. Sources, observations, claims, and contexts remain distinct. An Inquiry can accumulate evidence, disagreement, and rejected proposals. People accept research. Publication remains a later act.

hraness/kb is a different host and a different topic. It keeps Markdown and Git authoritative for agent memory. Indexes and graph views are replaceable. Search finds files. It does not become the record. Sponge applies that split to a research product: the run may spend tokens; the document trail stays inspectable.

Read Let the tokens flow on Sail Research, or through the hraness reading digest. This take is served on sponge.computer. The Context-1 take remains at A search subagent should return documents, not the answer. Agent locators are collected on developer resources.