Skip to content

README.md

Latest commit

 

History

History
54 lines (40 loc) · 4.35 KB

README.md

File metadata and controls

54 lines (40 loc) · 4.35 KB
Raw

Agents, Harnesses, and Skills

SEMOSS agents combine persistent conversations, reusable workspace configuration, model engines, executable tools, and skills. The default semoss harness runs the model/tool loop in Java and uses the same engine and project services as the rest of the platform.

Concepts

Term Meaning Lifetime
Harness Implementation of IAgentHarness that executes an agent request Registered runtime implementation
Agent / workspace A WORKSPACE project plus instructions, attachments, model defaults, and limits Reusable configuration
Room Model conversation with messages, options, and user ownership Persistent across turns
Run One invocation submitted through RunAgent Tracked by runId and status
Skill A SKILL project with SKILL.md and supporting files Reusable instructional package
Tool Executable operation exposed through built-in handlers or configured MCP resources Available according to run configuration and permissions
Execution target Room/insight, user assets, or project files the run works on Resolved for each run

An agent is not a model engine, a skill is not a tool, and the agent's workspace ID is not necessarily the ID of the project it edits.

Choose an execution path

Need Entry point
A persistent model turn whose tool results your client controls AskRoom, followed by AddToolExecution when needed
A server-managed agent loop with durable status and approval pauses RunAgent, normally with harnessType=["semoss"]
The Claude Code runtime through SEMOSS RunAgent with claude_code
A stateless model call LLM with history disabled; see model engines

Guide map

  1. Agent configuration: define a workspace, select a model and target, attach resources, and set limits.
  2. The SEMOSS harness: understand prompt composition, tools, compaction, hooks, and subagents.
  3. Run lifecycle and streaming: submit, monitor, pause, resume, and cancel runs.
  4. Agent skills: author, catalog, attach, stage, discover, and load skills.
  5. Workbench default agents: how frontend defaults and system agents are deployed and selected.
  6. PowerPoint visual inspection: a specialized agent workflow.

Requirements

Persistent rooms, workspaces, and agent run records use the model-inference database. It must be enabled and successfully initialized. Users also need access to the selected model, agent workspace, target project, and any resources invoked by tools.

The native harness uses a configured text-generation model. Tool use depends on the selected model's capabilities and provider integration. Python, Node, external harnesses, and specialized tools additionally depend on their deployment's runtimes and feature settings.

Platform agents and skills are shipped as project assets and cataloged at startup. ProjectWatcher, SystemDefaultEngines, and SystemAgentSeeder connect the distribution's project folders to the catalogs and workspace configuration. Deploy backend code and matching project assets together; see system agent deployment.

Source map

Area Source
Submission and registry RunAgentReactor, AgentHarnessRegistry
Context resolution AgentRunner, AgentConfigLoader
Native loop and tools SemossAgentHarness, HarnessToolExecutor
Runs and events Run services, stream services
Skill lifecycle Skill reactors and helpers
Built-in agent definitions SystemAgentSeeder, SystemDefaultEngines