Achilles Agents Library
AchillesAgentLib is the runtime layer through which the project executes skill-based agents, model-mediated calls, structured sessions, and family-specific subsystem behavior. The documentation is organized so that the shared runtime surfaces can be read first and the more local skill-family contracts can be approached afterwards.
The Common Runtime Layer
The most natural starting point is the common execution layer. RecursiveSkilledAgent is the top-level runtime entry surface through which discovery, registration, routing, and session-memory injection are coordinated. LLMAgent is the mediation layer for completions, interpretation helpers, output coercion, and session creation. Agentic Sessions describes the bounded multi-step execution model used by the runtime, together with the separate chapters for loop-based and SOP-based regimes. Subsystems and Skills explains why the library does not collapse all descriptor types into one misleading abstraction and how execution is divided across specialized subsystems.
These pages establish the shared contract of the system. Once that contract is clear, the family-specific pages become easier to read because they can be understood as local execution regimes built on top of the same agent runtime rather than as unrelated features.
Skill Families
The next layer is the family-specific contract. Orchestration Skills define bounded coordination over other skills and determine whether execution proceeds through loop sessions or SOP sessions. DBTable Skills formalize table-oriented workflows whose source specification in tskill.md is transformed into generated runtime code. Code Skills ultimately execute modules loaded from src/index.mjs or src/index.js. Dynamic Code Generation Skills follow a different contract, centered on runtime text-or-code decisions and temporary execution of generated snippets. MCP Skills govern MCP tool access through allowlists, scripts, and planning behavior. Anthropic Skills treat SKILL.md as the governing descriptor surface for a loop-session runtime that may expose scripts and resources.
Each family page explains not only what the descriptor looks like, but what the subsystem actually imports, prepares, generates, or executes at runtime. This is why the skill-family pages are separate from the general runtime pages: the common agent contract stays shared, but the local interpretation rules are materially different.
Supporting Components
The repository also contains supporting components that sit around the main runtime rather than replacing it. Backlog Manager is a small file-oriented utility for maintaining .backlog and .history task files. Evals Suite is the evaluation workspace under evalsSuite/ for measuring probabilistic runtime behavior, planning quality, intent detection, code generation, and model benchmarks. These components are important for operating and assessing the runtime, even though they are not themselves skill subsystems.