Planning, Search & Deep-Research Agents

What

Give agents structured planning (decompose, interleave, parallelize) and search (tree search over reasoning/actions) so they handle long horizons and hard tasks.

Papers

• LATS — Language Agent Tree Search (Zhou et al. 2023) — Monte Carlo Tree Search over both reasoning and actions, with value functions from self-feedback; beats ReAct and CoT on web/agent tasks.
• SPRINT — interleave planning and parallelized execution: plan chunks, fan out sub-executions, re-plan from results (great for long builds).
• ADaPT (Prasad et al. 2024) — decompose the task as needed: only plan deeper when the current step needs it (adaptive, cost-efficient).
• SWiRL — synthetic data generation + multi-step RL for reasoning-and-tool-use training.
• Search-o1 (2025) — large reasoning models that search the web agentically mid-reasoning to fill knowledge gaps; reasoning + retrieval fused.
• AlphaCode / AlphaCode 2 — competition-level code generation via massive sampling + filtering + (v2) a learned value model to select solutions.

Key ideas

• Search beats single-shot on hard problems; value models select winners.
• Adaptive decomposition saves cost (only expand what's needed).
• Retrieval fused into reasoning (Search-o1) beats RAG-after-the-fact.

SolOS mapping

• Search-enhanced Ask AI (adoption #3): Pulse chat + docs assistant get tool use — fetch live quotes/news/vault before answering (docs assistant already has list-pages/get-page; add web + quote tools).
• SPRINT-style parallelization: we already parallelize via delegate_task (up to 3 children) — formalize plan→fan-out→re-plan for multi-part builds.
• AlphaCode-style selection: our "sample-2-verify-ship" (note 01) is the same idea — verifier selects.