April 2026 Summaries

Sail Research has launched Sailboxes, a sandboxing platform designed for long-running AI agents that combines conventional sandbox APIs with durable-execution mechanisms intended to handle failures, preserve state, and support workloads lasting indefinitely. Sailboxes snapshot VM state, replay commands after failures with at-least-once semantics, offer checkpointing for nondeterministic operations, and proxy network traffic to support exactly-once handling of stateful network actions. The platform automatically pauses idle virtual machines and resumes them upon incoming traffic, charging users only for active CPU, memory, and disk usage rather than idle time, cold starts, or checkpointing. In a demonstration, four Sailboxes ran coding agents for a cumulative 110 hours to produce a Rust implementation compatible with Redis, consuming 64 active CPU hours and costing an estimated $3.65, compared with about $9.30 on Daytona, while several competing services were described as having runtime limits below the task duration. Sailboxes are generally available through Sail Research’s platform.

Sail reports that its agentic AI research system achieved 90.72% accuracy and 84.31% recall on the BrowseComp-Plus benchmark, approaching or exceeding cited closed-model results while claiming inference costs of about $0.15 per query, 6 to 35 times lower than several alternative providers. Its approach uses an orchestrator model, GLM-5.1, to formulate searches and reason over condensed findings, while a large parallel swarm of less expensive GPT-OSS-120B reader agents examines retrieved documents, filters irrelevant material, summarizes evidence, and requests additional context when needed. Sail argues that this division prevents the main agent’s context from becoming overloaded, allows broader retrieval with simpler tools such as Qwen3-Embed-8B and BM25, and makes token-intensive background research more economically feasible. The reported run processed roughly 6.47 billion tokens, with nearly all of them consumed by reader agents, illustrating the company’s view that efficient infrastructure for long-running, unattended agent workflows will be increasingly important as deep research systems scale.