liveBook · Manning

Code and tools layer

As explained in Self-Improving Agents

The deepest agent infrastructure layer, comprising the harness's source code, tool implementations, and orchestration logic.

The code and tools layer (L4) is the deepest agent infrastructure layer. It comprises the harness’s source code, tool implementations, and orchestration logic. Improving it can change more than code: it may also affect artifacts in other layers. section 1.3.2

One change, multiple layers

A change to L4 may regenerate a tool description in L1, restructure memory storage in L2, or rewire a metacognitive scaffold in L3. section 1.3.2

Figure 1.8 The harness artifact layers of self-improvement. The frozen model (LLM) serves as the unchanging substrate below. The four artifact layers stacked on it are shared infrastructure, inherited by every instance of the agent. Layer zero is the toolkit, the engineering work that measures, searches, and applies, reused at every layer.

Use the layer diagram to see L4 in relation to the other artifact layers, the frozen model, and the shared toolkit. section 1.3.2

How Self-Improving Agents covers it

How deep it goes
Identifies L4’s main artifacts and cross-layer effects, without detailing their implementations.
Examples
Uses changes to tool descriptions, memory storage, and metacognitive scaffolds to illustrate L4’s cross-layer effects.
What you can do after
Understand how harness infrastructure changes differ from updates to the frozen model’s weights.
How it connects
Connects the agent harness to prompts, memory, metacognitive scaffolds, tool descriptions, and the self-improvement loop.

This explainer was generated by Manning's knowledge graph from the text of Self-Improving Agents. The section links point to the parts of the book it draws on.