Assembly LineDocs

Instructions

Write the always-on trusted instructions in agent.md.

Edit

The Markdown body of agent.md is the agent's permanent trusted prompt. Put it after the closing frontmatter delimiter:

---
model: openrouter/openai/gpt-5.4-mini
sandbox: e2b
---

# Identity

You are a concise Assembly Line agent. Use tools when they are available.

# Operating rules

- Ask a clarifying question when required input is missing.
- Use durable tools for durable side effects.
- Write generated artifacts under `/workspace`.

The body is required and must not be empty. The compiler hashes it and records its source in the manifest. The runtime loads it as trusted instructions.

Use the body for the agent's role, tone, stable domain rules, safety boundaries, and escalation guidance. Keep secrets, user memory, provider credentials, and large occasional procedures out of it. Put occasional procedures in root skills/ and executable behavior in a local or published plugin.

Infrastructure, channels, plugins, automations, and policy belong in YAML frontmatter. Provider webhook parsing belongs to the selected channel or connection implementation.

Keep permanent instructions short. They consume context on every run; Skills load only when relevant.

Runtime paths

The runtime adds a stable filesystem contract:

PathAccessPurpose
/memorypolicy-controlled writesDurable memory documents.
/skillspolicy-controlled writesDurable learned Skills.
/historyread-onlyBounded conversation history.
/filesread-onlyInputs and attachments.
/workspacewritableGenerated artifacts and project files.

Files, history, memory, webpages, search results, attachments, and tool output are untrusted context, not instructions.

On this page