WorldSeed

The evidence shows role-filtered perception and a replayable event stream. The dedicated autoresearch runtime exposes only perceive/act tools and delegates filesystem work to the engine. The OpenClaw guide, however, allows agents to use their own file, shell, web, and other tools, so least privilege depends on the external runtime. Agents are encouraged to act autonomously; user confirmation is mainly available through optional intervention. Logs, pause/resume, and run records improve data-flow visibility, but retention, credential protection, and handling by external services are not fully described. Dependencies have version ranges but no security audit evidence is shown. Pause, stop, resume, and snapshots provide some recovery. The license and copyright holder are stated, but publisher identity is unverified and maintenance ownership or update path is unclear.

The protocol, architecture notes, and server code broadly agree on perception, action submission, and tick lifecycle. The sample describes OpenClaw, Codex, and a dedicated autoresearch runtime, so it does not establish that every execution path is fully consistent. The README lists Python, Node, uv, and LiteLLM prerequisites and setup steps, but broad version ranges and external model services leave availability dependent on outside components. The skill says failed actions return reasons and actual values, and the docs give health checks, state queries, and troubleshooting steps; the sample does not establish consistent handling across all error paths.

The materials explicitly target multi-agent simulations, production collaboration, games, and fictional worlds, with research, tool-pilot, layoffs, and espionage examples. Capability boundaries are described: WorldSeed manages shared state, visibility, and consequences, while agents' own tools do the work. Yet OpenClaw tools can be broad, so the boundary depends on configuration. Wake frequency, available actions, valid targets, and optional intervention are documented, while exact triggers largely depend on the scene and external runtime. Setup guides cover a local server, dashboard, OpenClaw, and Codex workflows, though the environment has several prerequisites and the sample does not cover every configuration.

Documentation is organized into getting started, architecture, scene configuration, runtime adapters, and development guides, including Chinese and English material. Setup instructions cover dependencies, commands, model credentials, and budget controls, but do not fully cover every deployment path. Names such as WorldSeed, worldseed, and Codex runner are mostly stable across the sample. There are numerous examples and troubleshooting steps, but no standalone FAQ or exhaustive common-failure guide. Some limitations, such as slow or offline agents and budget controls, are mentioned without a systematic limitations list. The repository includes the full MIT license text. The project version is 0.1.0, but no changelog appears in the sample. AIScientists Inc. is named as copyright holder, while maintenance responsibility, support commitments, and update process are unclear.

The product combines YAML scenes, role-specific perception, action rules, a dashboard, and run records to make multi-agent activity observable and replayable. The research example claims papers, experiments, and an evolution graph, but a static review cannot verify those outcomes. Information asymmetry, shared coordination state, and consequences provide a clear potential gain over using agents alone; the gain depends on scene design and external agent capability. Operation requires Python, Node, and a model service or runtime; research scenes may also consume GPU time and model credits. Budget controls are documented, but general cost estimates are not, limiting cost-benefit evidence.

Core architecture claims can be traced across the README, runtime docs, protocol structures, and server code, including perception data, action options, and event logging. This supports ordinary traceability and cross-source corroboration. Some demo metrics and experiment results appear only as project claims in the README and cannot be independently verified from this static sample. Test commands are documented but were not run. The materials distinguish system mechanics from demo narrative in places, but do not consistently label evidence for each result, so fact and inference separation remains limited.