Concepts
Core Concepts & Definitions
How documents preserve expertise, and how models, runtimes and authorization support a task together.
Glossary
- Document-first Agent
- Document-first design starts by preserving professional methods, materials and task instructions for a runtime to use within its authorized scope. Editing documents can change the approach; actual behavior also depends on the model, tools and execution permissions.
- Document-native Intelligence
- Prompts, rules, examples and long-term experience can be readable, editable documents. Documents help people understand and refine an agent's method; model capabilities, tool implementations and persisted execution state still come from their respective systems.
- Persistent Intelligence
- Documents preserve rules, knowledge and recorded progress for backup, comparison and migration. Changing models or environments requires checking dependencies, authorization and output quality. Continuing a complete installation also requires its runtime databases, configuration and credentials.
- Agent Harness
- The harness assembles documents and the current task into context, calls models and tools, and applies permission checks and task controls. Documents describe the intended method; the runtime carries it out in the actual environment.
- Environment Projection
- The environment includes the filesystem, tools, MCP, scheduler and external services. Documents provide task context and the environment supplies callable capabilities. Configuration and authorization determine what can be accessed and executed.
- Document Memory
- Useful facts, rules and progress can be organized into documents for inspection, revision and reuse. These serve a different purpose from conversation history, runtime databases and checkpoints. Backing up documents does not preserve all conversation and execution state.
An agent that organizes rules, knowledge and progress in readable documents.
A design approach that preserves reusable expertise in documents.
Useful experience that can be kept and resumed beyond a single conversation.
The runtime mechanism that organizes context, model calls and tool execution.
How an agent uses files, tools and external services to do practical work.
Long-term experience and work progress preserved as readable, editable documents.
Frequently Asked Questions
What is OpenJellyfish?+
OpenJellyfish is an open-source, document-first local Agent workspace that supports self-hosting. Use professional materials in everyday tasks, keep useful methods in documents, then configure a WeChat, web or API service for your team.
What is a document-first agent?+
It organizes professional rules, knowledge, examples and progress in readable documents so expertise can be reviewed, edited and reused. Execution also depends on models, tools, permissions and the environment. Editing a document does not grant new tool permissions.
How do documents relate to prompts?+
A prompt can be part of a document or assembled from documents and task context. Both can be version-controlled. Document-first design manages rules, knowledge, examples and work progress together so useful experience can accumulate.
How does the harness work?+
The harness organizes the task and document context, calls models, coordinates tool execution, checks permissions and maintains runtime state. Different models can use the same materials, but switching models still requires checking behavior and outputs.
What is the environment?+
The environment consists of the filesystem, tools, MCP, scheduler and external services needed for execution. Dependencies, access settings and credentials must be prepared in the deployment environment; copying documents alone does not supply them.
Does copying documents migrate the whole agent?+
Copying documents preserves rules, knowledge and recorded progress. For a complete migration, follow the deployment guide to back up configuration, runtime databases, the credential vault and encryption keys, then check dependencies and authorization in the new environment. Keep sensitive credentials out of public Git repositories.