coral-open-ended-discovery

name

coral-open-ended-discovery

description

Autonomous multi-agent open-ended discovery workflow inspired by the CORAL paper. Use when the task requires sustained search, iterative improvement, multi-agent exploration, shared knowledge accumulation, asynchronous parallel attempts, periodic reflection, or heartbeat-style redirection. Best for research discovery, algorithm design, systems optimization, skill discovery, long-running coding/search tasks, and open-ended problems where a single linear attempt is likely to plateau.

CORAL-Style Open-Ended Discovery

Overview

Use this skill when the user is not asking for a one-shot answer, but for continued discovery under uncertainty.

The core idea from CORAL is:

• run multiple long-horizon exploration threads
• accumulate reusable knowledge in shared memory/files
• periodically reflect and consolidate
• redirect when progress stalls

In OpenClaw, do not try to fully reproduce the CORAL infrastructure unless explicitly asked. Instead, map the paper's principles onto OpenClaw-native primitives:

• sessions_spawn → long-running parallel exploration
• memory/ + workspace files → shared persistent memory
• cron or heartbeat-style reminders → periodic reflection / redirection
• isolated sessions / subagents → asynchronous multi-agent organization

When To Use This Skill

Use this skill when the task has most of these properties:

1. Open-ended objective

   • “Keep improving this”
   • “Search for a better design”
   • “Explore multiple approaches”
   • “See what works best”

2. Long-horizon progress matters more than one-shot accuracy