Can General AI Agents Evolve from Data Gatherers to Professional Deliverables?

Manus is positioned as a generic AI agent that quickly integrates input enhancement (DeepSearch) and output generation (Artifact) while offering planning and reasoning capabilities, described by its developers as a "Less Structured" solution.

Observed Performance

Testing across several scenarios—travel planning, market attribution, intent‑driven suggestions, and artifact creation—revealed that Manus excels at information‑collection tasks but lags behind on deeper analytical work. Typical task durations range from 3 to 15 minutes, with code‑generation tasks (e.g., rewriting a pitch deck) completing in about three minutes.

Key observations include:

Longer processes often require human intervention for web access or robot verification.

User preferences are recorded, but the window for meaningful user correction is limited (3–10 minutes).

Intermediate artifacts such as todo.md and draft.md provide visibility into the workflow.

Agent Role Debate

The author questions whether a general agent should be a "collaborative agent" (type A) that works closely with users, or a "fully‑delegated agent" (type B) that delivers complete results autonomously. Drawing an analogy to private‑banking services, type A resembles advisory co‑decision‑making, while type B focuses solely on outcome delivery.

Current evidence suggests Manus behaves more like a fully‑delegated agent: it can produce a finished artifact within roughly 15 minutes, yet the quality of professional‑level deliverables remains lower than that of specialized analysts.

Root Causes of Limitations

The shortfall is attributed not to Manus itself but to two fundamental gaps in the reasoning pipeline:

Insufficient understanding and internalization of domain‑specific structures (the "Structure").

Inadequate high‑quality contextual data supplied as input.

Consequently, while the format of the output may be polished, its practical utility falls short of production standards.

Future Outlook

The author hypothesizes that, for a general agent to handle professional‑grade tasks in the medium to long term, it must incorporate private, structured intent and richer context. A collaborative model (type A) could mitigate current issues by allowing users to interrupt, correct, or supplement the agent’s plan, enabling a high‑level‑first, detail‑later workflow.

In summary, the evolution from a less‑structured, information‑centric agent toward a more open, collaborative system—capable of continuous human‑AI intent alignment—represents a promising direction for next‑generation general AI agents.