OpenAI4S, an open-source (MIT) scientific research agent, executes research actions as code cells in persistent Python and R kernels with an append-only Action Ledger, versioned artifacts, and workspace checkpoints for reproducibility. Across 36 research scenarios it scored 7.83 versus 5.7–6.4 for general-purpose coding harnesses, with the largest gains on long-horizon, compute-intensive workflows. The authors note full environment reproducibility remains unsolved for all evaluated systems, including their own.
An arXiv paper describes AfS (Agent for Science), a platform for long-horizon scientific work where projects run tens of hours across dozens of agent runs with only occasional human presence. Instead of trusting model behavior, the system makes non-compliance mechanically unrepresentable via enforced research laws spanning three time horizons. It targets a failure mode the authors say general coding agents inherit: under pressure to finish, they fabricate, skip, or smooth over.
An arXiv paper describes AfS (Agent for Science), a platform for long-horizon scientific work where projects run tens of hours across dozens of agent runs with only occasional human presence. Instead of trusting model behavior, the system makes non-compliance mechanically unrepresentable via enforced research laws spanning three time horizons. It targets a failure mode the authors say general coding agents inherit: under pressure to finish, they fabricate, skip, or smooth over.
A new arXiv paper introduces ScienceFlow, an LLM agent designed for long-horizon autonomous research. It addresses gaps in existing autoresearch agents by adding mechanisms for continuity, recovery from dead ends, and value-driven compute allocation. The goal is to sustain productive, stable, and goal-aligned research over extended periods.
A new arXiv paper introduces ScienceFlow, an LLM agent designed for long-horizon autonomous research. It addresses gaps in existing autoresearch agents by adding mechanisms for continuity, recovery from dead ends, and value-driven compute allocation. The goal is to sustain productive, stable, and goal-aligned research over extended periods.
Looking Glass is a local AI coding CLI with persistent, automatable sessions. Sessions survive terminal closure and a scheduler can trigger new AI turns later, reopening the workspace, retaining config, inspecting changes, and modifying code. It automates workflows like deploying and then checking results.
Looking Glass is a local AI coding CLI with persistent, automatable sessions. Sessions survive terminal closure and a scheduler can trigger new AI turns later, reopening the workspace, retaining config, inspecting changes, and modifying code. It automates workflows like deploying and then checking results.
An arXiv position paper argues reproducible research practices function as context engineering for AI coding agents. Agents lower the cost of maintaining tests, commit histories, repository structure, instructions, and decision records while making their benefits immediate. Researchers remain responsible for verifying these artifacts and the scientific judgments they encode.
An arXiv position paper argues reproducible research practices function as context engineering for AI coding agents. Agents lower the cost of maintaining tests, commit histories, repository structure, instructions, and decision records while making their benefits immediate. Researchers remain responsible for verifying these artifacts and the scientific judgments they encode.
ReproAgent is a four-stage pipeline that uses an implementation contract to guide AI agents in reproducing research papers as executable code. It addresses the challenge of split specifications in papers, improving reproducibility for AI practitioners.
ReproAgent is a four-stage pipeline that uses an implementation contract to guide AI agents in reproducing research papers as executable code. It addresses the challenge of split specifications in papers, improving reproducibility for AI practitioners.