Advanced concepts and architecture

For diagrams of the loop, harness, and archive, see Basic concepts.

Parent selection strategies

select_parent drives which generation is passed into the MetaAgent. Choosing the absolute best generation every time is actually flawed due to local maximums. HyperFlow provides five algorithms via select_parent.py:

• random: Pick any valid generation with equal probability. Max exploration.
• latest: Always pick the most recently created valid generation. Linear progression.
• best: Always pick the highest-scoring generation. Max exploitation.
• score_prop: Weighted random: higher scores get higher probability.
• score_child_prop (default): Score-weighted + child penalty. Encourages under-visited branches and exploring the global maximum.

Domains and Evaluation

A Domain defines what tasks the agent is evaluated on. - Evaluators: The grading strategy (evaluators.py). You can use static_evaluator (exact string match), llm_judge_evaluator (using LLMs for subjective questions), or human_feedback_evaluator. - Harness: The generic evaluation runner (harness.py). It formats the domain input, requests a forward pass on the TaskAgent, and evaluates it against the task context.

Architectures

Execution Modes

The environment is managed by Executor abstractions (executor.py): 1. LocalExecutor: Copies the repository to a localized temp directory. Faster, but has no sandboxing. 2. DockerExecutor: Creates a temporary Docker volume and sandboxes the process via Dockerode API (docker.py). Excellent for preventing arbitrary code execution escapes since the MetaAgent patches source code itself.

Self-Referential Improvement

A core concept from the HyperAgents paper is self-referential self-improvement ? the MetaAgent can modify its own instructions to become a better improver.

We achieve this with editable prompt files (using prompts_dir): - meta_agent.txt - task_agent.txt

If configured, the MetaAgent edits the very .txt files containing its internal constraints. Thus, the improver improves itself.

Early Termination

The loop contains intelligent optimizations: - Perfect Score Stop: Reaching 1.0 (100%) gracefully terminates further evaluation routines. - Score-Aware MetaAgent: The MetaAgent adjusts its strictness or focuses solely on edge-cases if the parent agent passes over 90% logic cleanly.