Glossary
Who This Is For
This page is for every author and reader of xAgent documentation.
What It Is
The glossary is the source of standardized terminology for xAgent documentation. Start with familiar AI terms, then explain their meaning and boundaries in xAgent. Product-specific terms identify interfaces, configuration, or implementation details rather than replacing general concepts.
xAgent is an agentic platform for ongoing work. Agentic capabilities are the core; private deployment, multi-user access, and centralized configuration are deployment and administration extensions.
Core AI Terms
| Term | Standard Meaning |
|---|---|
| Agent Harness | Runtime mechanisms around the model that organize tools, context, memory, progress recovery, collaboration, and human review; not a model, a prompt, or a certification standard |
| Agent | An entity that understands goals, plans, calls tools, and progresses work based on execution results |
| Agentic Loop | A cycle of model decisions, tool execution, result feedback, and further decisions; it does not guarantee task success |
| Tool Calling | Performing concrete operations through tool interfaces and using their results to inform the next step |
| Tool | A concrete capability an Agent can invoke, such as reading a file, querying data, or sending a message |
| Skill | A reusable task method, constraints, and tool-use guidance; not an individual tool or an automatic grant of tool permissions |
| Context Engineering | Organizing instructions, materials, history, and capability definitions for the current task; xAgent includes on-demand Skill and tool loading and context compaction |
| Context Compaction | Reducing session history while retaining goals, constraints, progress, and key results for continued execution; not long-term memory, infinite context, or lossless retention |
| Agent Memory / Long-term Memory | A standard feature enabled by default that reuses stable preferences, accepted decisions, and background across one user's sessions, checked against source evidence, time, and current facts; not storage of every conversation |
| Resumable Execution / Tasks | Continuing tasks after waiting or interruption using session state, recovery snapshots, and work artifacts; not a guarantee of exactly-once external actions |
| Agent Teams | Multiple agents organized through dedicated managers, member session trees, role reuse, manager delegation, owner guidance, and shared materials; the product's Virtual Teams, not multi-user account administration |
| Multi-agent Orchestration | Organization of agent roles, task delegation, and collaboration; xAgent executes through teams and sessions, while WorkGroup stores orchestration configuration. Publishing configuration is not task completion |
| Multi-agent Collaboration | Independent Agent sessions collaborating through explicit tasks, messages, and artifact references, including parent-child delegation; not shared full context or cross-user authorization |
| MCP (Model Context Protocol) | A protocol for connecting external tools and capabilities; usability still depends on connection state, configuration, and authorization |
| A2A Client (Agent2Agent Client) | User-scoped connections to remote Agents, remote task tracking, and result delivery; current client support does not mean xAgent exposes an A2A server |
| Human-in-the-loop (HITL) | Human input, review, and action approval within a task; approval coverage depends on the tools and policies involved |
| Observability | Inspecting execution through task status, messages, tool results, and artifacts; not a full automated evaluation platform |
| BM25 Retrieval | Term-relevance retrieval for memory and artifacts; it must not be described as vector retrieval or GraphRAG |
Product and Technical Terms
| Term | Standard Meaning |
|---|---|
| Task | A goal unit executed by an Agent |
| SubAgent | An Agent session that accepts a defined subtask, maintains its own context, and executes work independently |
| Session | A context for user-Agent interaction, task execution, and retained state |
| Workspace | The space for files, context, and intermediate artifacts used by an Agent during a Task |
| AgentPlugin | A plugin service connecting external systems, user Channels, and capability declarations; called Connector before 0.0.16.beta |
| ProcessSandbox | The platform isolation boundary for untrusted processes, responsible for file mounts, environment, process trees, and resource limits |
| Runtime Assets | Execution dependencies independently installed and verified by xAgent, then provided read-only to ProcessSandbox |
| Execution Lease | The minimal file view, overlapping-write coordination, and file-change commit capability that WorkspaceFileService provides for one process execution |
| Trigger | A mechanism that turns a future time or external event into a Session Task |
| Virtual team | Product name for Agent Teams; Server 0.1.0 provides dedicated managers, member session trees, role reuse, manager delegation, owner guidance, shared materials, and whole-team archiving |
| WorkGroup | Implemented orchestration configuration for Agents and communication links, with drafts and published versions; Sessions execute the actual tasks |
| Work Record | Session-scoped topic records, revisions, and deltas used to inspect continuity and evidence |
| Model Profile ID | Stable reference to a model configuration; distinct from its display name or the Provider's model ID |
| Delivery Receipt | Evidence that a particular messaging stage was accepted; does not by itself prove the whole task completed |
When to Use It
User guides, technical references, and maintenance documentation should all follow this glossary. In user-facing pages, explain the work value first, introduce familiar terminology when needed, and then use it consistently without piling up acronyms. Preserve exact UI labels in operating instructions.
Writing Rules
When writing documentation:
- Start with what the user can accomplish, then explain the Agentic capability; do not replace product positioning with implementation names.
- Do not invent synonyms or stretch feature claims to fit popular terminology.
- Do not describe planned capabilities as implemented; identify experimental features, version differences, and applicable boundaries.
- Distinguish memory, context compaction, and information retrieval; distinguish successful calls, message delivery, and task completion.
- Update this page, its translation, and related concept pages when adding a term.
- Explain how to use a feature in user-facing pages; reserve internal boundaries for technical references.