Neura Market
prompt logo

prompt

Free

Design self-referential meta-agents with a unified program

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About prompt

The HyperAgents Designer prompt provides a comprehensive design specification for creating self-referential meta-agents, where the task layer and meta layer coexist within a single editable program. It outlines core responsibilities such as defining a unified program with task policy, meta policy, and versioned history; specifying a self-modification interface with immutable regions and edit operators; grounding every edit in evidence with regression testing; bounding recursion with caps and mandatory cool-downs; and preserving auditability with signed diffs and append-only traces. The prompt is based on Meta FAIR's HyperAgents research (arXiv 2603.19461) and is part of the ai-boost/awesome-prompts collection on GitHub.

Key Features

Unified single-program architecture combining task policy, meta policy, and versioned history with rollback support
Self-modification interface with allowed reads/writes, immutable regions (safety rules, eval harness, kill switch), and edit operators (replace, refine, add-tool, prune-tool, swap-prompt)
Evidence-based editing requiring failing trajectory, hypothesized cause, expected metric improvement, and regression suite pass
Bounded recursion with caps on edits per cycle, artifact size, and meta-on-meta depth, plus mandatory cool-down runs between edits
Auditability via signed diffs with written rationale, append-only traces and evaluations externally readable
Hard kill switch reachable from outside the program and external human approval for edits touching meta layer, eval harness, or safety constraints

Pros & Cons

Pros
  • Comprehensive and rigorous design specification based on academic research
  • Clearly defined safety and bounded recursion mechanisms to prevent runaway self-modification
  • Open-source and freely available as a prompt for AI tools
  • Includes versioned history and audit trails for transparency
Cons
  • Requires deep understanding of meta-agent concepts and programming
  • No ready-made implementation; only a design prompt
  • Self-modification may introduce unpredictable behavior if safety bounds are not carefully enforced

Best For

Designing self-improving AI agents that rewrite their own source codeImplementing meta-learning systems with bounded self-modificationResearch in agent architectures for autonomous improvementBuilding educational prototypes for exploring recursive self-improvement in AI

FAQ

What is a HyperAgent?
A HyperAgent is a self-referential meta-agent: a single editable program where the task layer (solves the user's task) and the meta layer (edits the task layer) co-exist and can rewrite each other under bounded supervision. Improvements compound across runs because the agent's own definition is its working memory.
How does self-modification work?
The meta layer reads the full source, traces, and evaluations, and writes to specified sections using edit operators (replace, refine, add-tool, prune-tool, swap-prompt). Certain regions like safety rules and the kill switch are immutable. Every edit must be grounded in evidence (a failing trajectory, hypothesized cause, expected metric improvement) and pass a regression suite.
What prevents runaway self-improvement?
Bounded recursion is enforced with caps on edits per cycle, edits per artifact, and depth of meta-on-meta. There is a mandatory cool-down of N task instances between consecutive edits, and a hard kill switch reachable from outside the program. External human approval is required for any edit that touches the meta layer itself, the eval harness, or safety constraints.