Introduction

HyperAgents is a self-improving agent framework. Instead of hand-tuning prompts and tools, you run an evolutionary loop: a MetaAgent reads evaluation results and edits the TaskAgent’s codebase (prompts, domain logic, tools); each version is scored and stored in an archive so future generations can branch from the best ancestors.

The design is inspired by HyperAgents (Meta Research, 2026) and ported to TypeScript with a generic, pluggable layout: LangChain for models, LangGraph for the tool loop, and a small core for evolution, archives, and execution.

What you get

  • TaskAgent — solves tasks for your domain (optional tools, ReAct-style loop).
  • MetaAgent — uses built-in bash and editor tools to patch the repo under evaluation.
  • runGenerateLoop — selects parents, applies lineage patches, runs MetaAgent then evaluation, updates JSONL archive.
  • Domains — implement loadTasks, evaluate, formatInput, report for your tasks.
  • Executors — local (fast) or Docker (isolated) workspaces per generation.

Mental model

Think of the MetaAgent as the teacher updating the textbook, and the TaskAgent as the student taking exams with the latest textbook. The archive is the gradebook plus revision history.

For narrative and workflow diagrams together, see Basic concepts.

Relation to the research paper

The paper’s idea of self-referential improvement — the improver changing its own instructions — maps to editable prompt files (promptFile / promptsDir) in the workspace. The underlying LLM weights are not trained; all improvement is external (code and prompts). See Limitations.

Next steps