Manus AI agent: multi-agent architecture and the wrapper debate

Executive overview

Most AI agent products are "wrappers" — combining existing models with tool integrations. The debate is whether that's a valid approach or a dead end. Manus shows wrappers can deliver real value through thoughtful architecture, transparency, and targeted fine-tuning.

A wrapper's competitive durability depends entirely on what it builds that competitors can't easily replicate.

How Manus works

• A planner agent breaks the user's prompt into a master plan of subtasks before any execution begins
• Subtasks are handed to specialised sub-agents, each with its own domain (memory, knowledge, execution)
• Sub-agents draw on 29 integrated tools — web navigation, code execution, file parsing
• An executor agent synthesises all outputs into a final result
• Chain-of-Thought Injection keeps agents stable across dozens of reasoning rounds by forcing active plan reflection

Benchmark performance

• Humans score ~92% on the Gaia benchmark; OpenAI Deep Research peaks at 74%
• Manus scored 86.5% — the highest any AI agent has achieved on Gaia
• Gaia tests reasoning, multimodal handling, web browsing, and tool proficiency

The wrapper question

• Manus uses Claude 3.7 Sonnet as its core model, plus open-source tools (browser-use, E2B sandbox)
• Cursor, Harvey, and Windsurf are also wrappers — the category label doesn't determine quality
• Manus co-founder's strategy: work orthogonally to model development, so each new model release is an opportunity, not a threat
• What separates good wrappers: intuitive UI, proprietary evals, careful fine-tuning, well-designed multi-agent architecture

Strengths and limitations

• Per-task cost ~$2 — significantly lower than Deep Research
• Exposes the file system so users can watch agents work in real time; competitors are largely opaque
• Users can inspect, customise, or replace individual sub-agents and tool integrations
• Coordination breaks down as task complexity scales
• Core advantages (UX, fine-tuning, integrations) are replicable by well-resourced competitors
• Vulnerable to API pricing changes or provider policy shifts

What durable differentiation looks like

• Invest early in proprietary evals that are expensive or slow to replicate
• Embed workflows deeply into specific user routines to raise switching costs
• Secure integrations with platforms or datasets competitors can't easily access