> For the complete documentation index, see [llms.txt](https://docs.darwinslab.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.darwinslab.ai/the-synthetic-darwin-protocol/agent-structure.md).

# Agent Structure

Agents are lightweight, containerised modules with well-defined state and mutation potential.

| Component                     | Purpose                                                                                                                                                                                                                                                                                                                       |
| ----------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Base Model**                | Pluggable - selected at run-time from a model registry that ships with presets for OpenAI O-series (O3-mini / O3-high / GPT‑4o), Anthropic Claude 3.x (Sonnet / Opus), Meta Llama-3 (405 B / 70 B), DeepSeek Coder / Reasoner, or any open-source checkpoint addressable through an OpenAI-style or Anthropic-style endpoint. |
| **Task Strategy**             | Blueprint describing how the agent designs, mutates, and orchestrates children.                                                                                                                                                                                                                                               |
| **Mutation Logic**            | Parameter-level rules (e.g., temperature, context window) plus structural edits (swap attention heads, change prompt templates, introduce new reasoning chains).                                                                                                                                                              |
| **Memory & Cache**            | Short-term scratchpad + long-term vector store for episodic recall.                                                                                                                                                                                                                                                           |
| **Fitness Prediction Module** | Local heuristic that estimates expected reward before expensive evaluation runs.                                                                                                                                                                                                                                              |

*An adapter layer automatically normalizes provider-specific chat formats (function-calling JSON, Claude tool-use, etc.), so adding/swapping a model remains a one-line YAML change; no code edits are required. Multiple base-model variants can be instantiated in parallel; the adapter layer merely unifies their interfaces.*
