Quiz: Agent System — OpenClaw Academy

1 What are the four core components that make up an OpenClaw agent?

A model provider, a webhook URL, a database connection, and a Docker container An LLM, tools, a workspace, and sessions (plus skills) A REST API, a message queue, a session token, and a config file A Telegram bot, a shell script, a cron job, and a log file

2 Which workspace file controls the agent's personality, tone, and behavioral boundaries?

AGENTS.md — because it contains the agent's operating instructions TOOLS.md — because it tells the agent which tools to use SOUL.md — because it defines persona, voice, and what the agent won't do IDENTITY.md — because it contains the agent's name and emoji

3 What does TOOLS.md control?

Which tools the agent is allowed to use (the tool allowlist) Tool execution timeouts and sandboxing behavior Nothing — it's guidance notes for the agent, not a policy file The list of skills loaded at startup

4 In the agent loop, what happens when the LLM responds with 'stop_reason: tool_use'?

The loop ends and the tool call is returned to the user as the final response The requested tool is executed, the result is appended to the conversation, and another LLM call is made The agent pauses and waits for user approval before running the tool The Gateway raises an error because tool_use is not a valid stop reason

5 What is the difference between session compaction and context pruning?

Compaction deletes old sessions; pruning deletes old tool results Compaction summarizes old conversation history (modifies JSONL on disk); pruning trims oversized tool results from in-memory context only (disk unchanged) Compaction and pruning are the same thing — different names for the same process Compaction is manual (you trigger it); pruning is automatic

6 In a multi-agent setup with bindings, if both a peer-level binding and a channel-level binding match an incoming message, which wins?

The channel-level binding wins because it's more general and covers more cases The last binding in the list always wins regardless of specificity The peer-level binding wins because more-specific matches always take priority Both agents receive the message (broadcast behavior)

7 What does the 'memory flush' feature do before session compaction?

It clears the MEMORY.md file to prevent stale information from being re-injected It runs a silent agent turn that prompts the agent to write important context to a daily memory log before history gets summarized It uploads the session transcript to a cloud backup service It sends the user a notification that their session is about to be compacted

📝 Module 4 Quiz: Agent System