Manus Modules System

Overview (confidence: high)

Manus uses a wide XML envelope with 15+ module blocks, making it the most verbose prompt in the corpus at approximately 30K+ tokens. It is a multi-file agent system with an agent loop pattern and module-based orchestration.

Envelope Format (confidence: high)

Wide XML envelope — 15+ <module> blocks, each handling a different capability or domain. This is the most extreme example of the xml-envelope-format in the corpus in terms of breadth, though it pushes against the scaling limits of monolithic prompts.

Agent Loop Pattern (confidence: high)

Follows a plan → execute → observe → repeat loop. The agent plans its approach, executes tools, observes results, and iterates. Tasks are tracked through a task-state-machine lifecycle.

Module-Based Orchestration (confidence: high)

Each module has specific dependencies that are activated per task. The system activates only the relevant modules for a given task, enabling specialized behavior without loading all 15+ modules simultaneously. The modules cover diverse capabilities from coding to web browsing to data analysis.

Additional Patterns (confidence: high)

• Proactive memory system: Reads from and writes to persistent memory between sessions
• Skills directory gate: Routes tasks to appropriate modules
• Exhaustive enumeration: Lists all options without truncation

Implications (confidence: high)

Manus’s module system is the most scalable single-agent architecture in the corpus. The 15+ module approach enables specialized behavior per task type but pushes monolithic prompts to their limit — 30K+ tokens makes maintenance and debugging challenging.

Sources

• [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]