Vibe Coding
Back to FastClaw

FastClaw / Run the model

agent

Agents are the core objects of FastClaw: models, system files, memories, skills, tools, sessions, sandboxes, and channel bindings are all organized around Agents.

View official sourceDocs index
agent key concepts infographic
agent key concepts infographic

Intelligent body factory

The README refers to FastClaw as Agent Factory. Each Agent can have independent provider/model, system prompts, skills, tool switches, workspace, memory, channels and sessions.

fastclaw agents init alpha \
  --provider openai \
  --model openai/gpt-4o-mini \
  --api-key-env OPENAI_API_KEY

fastclaw agents config alpha set temperature 0.7
fastclaw agents config alpha set sandbox.enabled true

System files

Agent system files are the core materials of prompt and memory layers. System files allowed by source code include SOUL.md, IDENTITY.md, USER.md, BOOTSTRAP.md, MEMORY.md, HEARTBEAT.md, AGENTS.md, TOOLS.md, agent.json.

Filefunction
SOUL.mdThe Agent's long-term personality, goals, and work style
IDENTITY.mdIdentity overlay, takes precedence over regular context
USER.mdCurrent chatter or upstream user information
MEMORY.mdlong-term facts and preferences
AGENTS.md / TOOLS.mdProject rules and tool descriptions

Prompt assembly

Agent's system prompt will be assembled according to modules: SOUL/IDENTITY, memory, confidentiality, sandbox, task delegation, skills, group chat, thinking, tool discipline, workspace update, etc.

  • Prompt mode supports agent, chatbot, and customize.
  • IDENTITY OVERRIDE is placed in a key position to ensure that identity documents take priority.
  • bindSession will bind session/project/coding root/runtime to the current Agent loop.