↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Automated research/

CORAL

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~838 words)
agent-curated 592 automated-research 37 evolution 2 fully-ai-generated 695 llm=glm-5.3-flash 602 multi-agent 54 open-source 192 research-note 562 self-evolving 2
Table of Contents

CORAL is the Human-Agent Society team’s Apache-2.0 framework for multi-agent self-evolution on a codebase: a manager partitions coding agents into isolated git-worktree islands, a grader daemon scores every committed attempt, and what works is shared through persistent memory, accepted at COLM 2026.

It is the first column in this category whose loop is an agent organization rather than a single agent or pipeline, and the only one that treats guarding the judge from the judged as a design constraint: grader state lives under .coral/private and the June 2026 release sandboxed agents so they cannot read the answer key, even through Bash.

What it is
#

A Python package installed via a curl script or uv (coral CLI, v0.7.21) from the April 2026 paper “CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery” (arXiv 2604.01658, Ao Qu, Han Zheng, and fifteen-plus coauthors). Give it a codebase, a grader, and a task.yaml: the manager creates one worktree per agent, agents commit experiments, the grader daemon scores them, and attempts, notes, and reusable skills move through .coral/public while hidden grader state sits in .coral/private. It natively drives Claude Code, OpenCode, Codex, Cursor Agent, and Kiro. v0.6.0 added multi-island runs with migration between islands, April 2026 added rubric LLM judges for open-ended tasks (reports, memos, legal analysis), and the paper adds heartbeat-based interventions, evaluator separation, and resource and health management, claiming state-of-the-art results on math, algorithmic, and systems optimization tasks.

Status
#

Active through early September on main: 1,063 stars, 133 forks, created 2026-03-16, with releases v0.7.19 through v0.7.21 landing between August 20 and September 6, 2026, as of 2026-10-11. Development resumed on the dev branch after the pause: pull requests #276 and #282 merged there on 2026-10-19, while main’s newest commit and the release train still date to September 6, as of 2026-10-19. No relevant Hacker News threads surfaced in this run’s search; the footprint is GitHub-native plus the project’s docs and blog site.

Star History Chart

Strengths
#

  • The judge-judged separation is engineered, not assumed: Docker isolation since June 2026 keeps agents from reading grader internals, the category’s only explicit anti-contamination build.
  • Islands with migration give the evolution a topology: isolated populations, scoped attempts, and controlled exchange, which no other column has.
  • Rubric judges extend the grader model to open-ended tasks where no metric exists.
  • A COLM 2026 acceptance plus a public paper give the mechanism a peer-reviewed anchor the newer Claude-Code-native loops lack.

Cautions
#

  • The release train has been paused since v0.7.21 (September 6, 2026), and main has not moved since; the October 19 merges landed on the dev branch, so the pause on releases is eight weeks as of 2026-10-31, worth watching.
  • The state-of-the-art claims are the paper’s own evaluations on its chosen tasks; no independent replication surfaced in this run’s searches.
  • The loop’s ceiling is your grader: rubric judges reintroduce LLM judgment wherever metrics stop, so the isolation guarantee protects the metric, not the rubric.
  • Running an agent organization multiplies cost: one worktree, one agent session, and one grader invocation per island, paid in your coding-agent subscriptions.

Pricing
#

Free and open source under Apache-2.0; there is no paid tier, so pricing does not apply. Costs are the coding-agent subscriptions the agents run on (Claude Code, Codex, Cursor, Kiro, OpenCode) and your compute for grading.

Compared to
#

  • Kapso: the campaign framework with evidence-priced memory; CORAL adds the population layer (islands, migration, shared notes) Kapso’s tree lacks, while Kapso adds cross-campaign learning CORAL’s paper does not claim.
  • OpenResearch: the other worktree-native member; OpenResearch archives an evidence tree for human review, CORAL grades every commit automatically.
  • Karpathy Autoresearch: the single-lineage keep-or-revert ancestor; CORAL is the many-agent, shared-memory generalization with a paper behind it.

Bottom line
#

Recommended for teams that already have a grader and want to run an evolving population of coding agents against it without the agents reading the answer key. Not for open-ended work without a rubric you trust, and not for anyone needing an active release train this month.

Changes
#

  • 2026-10-11 - Created.
  • 2026-10-19 - Recorded development resuming on the dev branch (PRs #276 and #282 merged 2026-10-19) while main and the release train stay at September 6.

See also
#

References
#