planning

CEO-Bench introduces a simulation benchmark that evaluates language model agents' ability to manage a startup over 500 days, testing long-term planning, noise handling, adaptability, and multi-task coordination. Results show that even the strongest models struggle, with only Claude Opus 4.8 and GPT-5.5 finishing above the starting balance.

Matt Pocock introduces a decision-mapping skill to split planning into multiple sessions, similar to /to-issues, aiming to streamline greenfield and brownfield builds.

The author built a personal AI agent that uses a frontier model (Codex) for high-level planning while running most token processing locally on a dual RTX 3090 system, enabling long-duration tasks with deterministic validation. The agent supports three swappable tiers: planner, local, and senior, and is available as an open-source repository.

COMET is a model-based reinforcement learning algorithm that combines a frozen object-centric encoder with a transformer-based world model and Monte Carlo Tree Search, using causal attention to focus on task-relevant objects, achieving higher scores on visual RL benchmarks.

Deep Work Plan is a product that helps users provide their AI agents with a structured plan, emphasizing the importance of context over models.

A developer shares satisfaction with Opus 4.8 for planning and GPT-5.5 for execution, emphasizing that breaking tasks into smaller steps improves quality and that dynamic workflows are underrated.

The paper proposes SVoT, a reinforcement learning framework that generates interleaved, verifiable intermediate states and visualizations for multi-hop spatial reasoning in MLLMs, achieving significant accuracy gains on new benchmarks involving multi-object interactions and numerical reasoning.

A discussion about deploying multi-agent AI systems in production, where different agents handle planning, execution, communication, and project management, asking about real-world experiences and bottlenecks.

This paper introduces PhysTool-Bench, a benchmark for evaluating multimodal large language models' ability to recognize and plan the use of physical tools in real-world scenes. The authors find that even the best model identifies only 58.7% of tools and completes just 21.0% of queries end-to-end, revealing a two-level deficit in perception and functional commonsense.

Introduces front-to-attractors (F2A), a new heuristic class for bidirectional search that reduces computational cost by evaluating distances to a small set of attractors instead of the full opposite frontier, achieving up to 11.2x fewer pairwise evaluations and 4.8x fewer node expansions than existing methods.

This paper systematically reviews text world models for LLM-based agents, covering foundations, construction paradigms, applications in planning and training, and evaluation methods.

Stride is an AI-powered workspace that assists with planning, designing, and shipping projects.

A practitioner observes that limiting AI agents to plan only one step ahead instead of multiple steps significantly improves reliability in real-world automation workflows involving CRM and lead qualification, as long-range plans become brittle when external state changes.

This paper introduces OCLGen, a compute-efficient test-time search algorithm that integrates generative planning models with a classical Open-Closed List framework, improving solution quality across combinatorial planning domains.

A comprehensive survey of world models that provides a multi-axis taxonomy covering architectures, methodologies, reasoning strategies, and applications across AI domains, including key systems like Dreamer, MuZero, and Sora.

A developer shares their workflow using Cursor's subagent harness with Opus 4.8 Max Thinking for long context understanding and implementing large features in Swift, emphasizing hands-on planning and phased acceptance testing.

DecomposeR introduces a planner-centric reinforcement learning framework that represents research plans as typed DAGs, enabling finer-grained optimization of planning and execution for deep research tasks, achieving 5.1–8.0 point improvements over open baselines.

This paper proposes three rerooter designs for Levin Tree Search that leverage state-space structure and learned heuristics to improve search efficiency without explicit subgoal generation, achieving state-of-the-art online training efficiency.

This paper investigates how to encode factored planning tasks (FTS) into SAT, proposing multiple encoding strategies and analyzing the impact of task transformations on SAT-based planning performance. It aims to extend SAT solving to more compact planning representations beyond heuristic search.

This article discusses an anti-pattern in AI agent systems where agents appear busy but fail to complete tasks. The author suggests separating responsibilities and requiring proof of completion as a solution.