Tag
This paper introduces Transfiver, an architecture for human-AI co-inference through a shared editable state, enabling direct updates by both humans and models to improve long-term interactions.
An experiment where thirteen AI models from different providers interact in a shared world via cron scheduling, leading to emergent cross-references and collaborative idea-building without explicit design.
This paper introduces MASS, a method for multiplayer world models that disentangles world dynamics from view rendering using an authoritative shared state, enabling scalable and consistent multi-agent simulation with up to 1,024 concurrent players.
The article discusses three unexpected problems when running multiple coding agents in parallel: conflicts over shared working tree, runtime collisions (database, ports), and difficulty detecting stuck agents. Solutions include using per-agent git worktrees, isolated runtimes, and monitoring remaining gap metrics.
Discusses the failure modes of shared state in large multi-agent systems with 50+ nodes, including race conditions, multi-node desync, and poisoned context, and asks for war stories from practitioners.
Lemma is an open-source shared workspace that integrates humans and AI agents into a team, collectively reading and writing spreadsheets, files, workflows, and other states, reducing glue code and scattered work in chat logs.
Introduces PatchOptic, an interface for shared-state LLM workflows that uses projected reads and verified structured patches to ensure valid updates. Evaluated with PatchBench across 46 cases, showing reduced token cost and leakage while maintaining quality.
Azalia Mirhoseini highlights DeLM, a decentralized language model approach where agents communicate via shared state, achieving ~10% improvement on SWE-bench Verified with Gemini-3 Flash at less than half the cost.
The article discusses the problem of stale context in AI agent systems, where agents make decisions based on outdated information, and proposes a coordination primitive with versioning and presence signals to prevent conflicts and wasted tokens.
This paper studies failure modes in shared-state collaborative reasoning for resource-constrained visual agents, introducing CoSee, an auditing framework that formalizes read-write-verify loops. It finds that naive shared workspaces can amplify hallucinations and identifies noise reinforcement and policy collapse as dominant failure modes.