Why Modern AI Agents Are Getting Lighter, Thinner, and More Collaborative

The piece revisits the concept of "harness engineering" for AI agents, noting that since the emergence of agents the need for a complete control system—akin to reins, saddle, and protective gear—has persisted, but OpenClaw has shifted AI sovereignty from model vendors to the user side.

Three Representative Agents and Their Design Philosophies

Manus : A turnkey black‑box solution where the second (ability) layer—memory, system prompts, knowledge base, workflow, MCP— is fully managed by Manus. Users receive only a finished UI, comparable to buying a fully assembled car and only operating the accelerator and steering wheel.

OpenClaw : Provides an open skeleton; the ability layer is decomposed into a set of text protocols (agent.md, soul.md, user.md) plus heartbeat, skills, and session management. All these assets belong to the user and can be tuned via natural‑language prompts. The UI expands beyond a browser to IM platforms such as Discord, Feishu, and DingTalk, similar to a customizable chassis where the engine is supplied but the suspension and aerodynamics are user‑installed.

Claude Managed Agent : Compresses the ability layer to three core primitives—Environment, Session, and Events—while omitting explicit memory, knowledge base, or workflow modules. These capabilities are “sunk” into the model itself, allowing the agent to plan, reason, and act within a minimal framework.

Trend: Agents Are Getting Lighter and Architectures Thinner

As models become more capable, the heavy "thick agent" approach (extensive middle‑layer tooling) is giving way to slimmer designs that rely on the model’s intrinsic abilities, reducing engineering overhead while preserving functionality.

Multi‑Agent Collaboration Is Getting Thicker

Complex, long‑running tasks are prompting a shift toward multi‑agent teams. Three common collaboration paradigms are described:

Manager‑Worker : A central manager splits a task into independent sub‑tasks for workers, then aggregates results. Ideal for tasks that can be cleanly partitioned, such as industry competitor analysis or software component design.

Sequential / Pipeline : Agents operate in a linear chain where each stage depends on the previous output (e.g., data cleaning → feature engineering → model inference → visualization). This offers clear traceability but struggles with back‑tracking or iterative revisions.

Peer‑to‑Peer / Decentralized : No central controller; agents negotiate, broadcast, and share information autonomously. Suited for emergent scenarios like social simulations, multi‑party negotiations, or self‑organizing load balancing, though it introduces higher coordination complexity.

Examples include the HiClaw project’s 7‑million‑car design and the AI Hedge Fund initiative, which assemble a legion of specialized agents (e.g., Buffett‑style value investor, Munger‑style contrarian) to generate collective insights that surpass any single agent’s capability.

Overall, the evolution mirrors the internet’s trajectory: devices become lighter, yet collaborative networks grow richer, unlocking exponential group intelligence.