Limitations

While HyperAgents enables powerful self-improving agents, several technical and practical limitations exist in the current implementation.

Frozen foundation models

The most significant limitation is that the underlying Large Language Model (LLM) is frozen.

  • HyperAgents improves the Agent Logic (TypeScript code, prompts, and tools), but it cannot update the weights or training of the model itself.
  • All improvements are "external" to the model's core architecture.

Framework immutability (self-update paradox)

The Agent cannot evolve its own core classes!

Because HyperAgents is often installed as a library (via npm/pnpm), the core code for TaskAgent and MetaAgent lives in node_modules/.

  • The MetaAgent cannot modify its own framework code or its class definitions.
  • Even in a local setup, the evolutionary loop targets a specific repoPath. If the agent's primary logic is outside that directory, the MetaAgent cannot reach it.
  • This is why the framework focuses on evolving prompts in external files and local domain logic — the core agent infrastructure remains static and generic.
  • Workaround: use promptsDir to place editable prompt files in the workspace so the MetaAgent can modify its own instructions.

Fixed task distribution

Currently, the system optimizes for a fixed set of tasks provided by the user. Truly unbounded open-endedness requires a system that can also generate its own tasks and curriculum (co-evolution), which is a future research direction.

Static outer loop

While the agent can modify almost all of its own codebase, the outer evolutionary process is currently human-engineered:

  • Parent Selection: The logic for choosing which generations to branch from is fixed.
  • Evaluation Protocols: The way task scripts are run and scored is defined in the initial configuration.

Computational cost

Evolutionary computation is resource-intensive.

  • Generations: Each experiment typically requires 10–50+ iterations for meaningful improvement.
  • Costs: Each generation makes multiple LLM API calls (MetaAgent + TaskAgent per task). A full run with GPT-4o can cost $5–50+ depending on the number of generations and tasks.
  • Latency: Each generation takes 20–120 seconds depending on task count and model speed.

For what the framework can do well, see Introduction and Basic concepts.