AI Coding Agents · Human-Agent-Society

CORAL

CORAL is a Python framework for orchestrating multiple AI coding agents (Claude Code, Codex, Cursor, etc.) to autonomously solve problems through iterative improvement. Agents work in parallel git worktrees, share findings via a common state directory, and are evaluated by a pluggable grader daemon.

Talk to DEV.co Discuss Implementation

Source: GitHub — github.com/Human-Agent-Society/CORAL

783

GitHub stars

101

Forks

Python

Primary language

Apache-2.0

License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

Field	Value
Repository	Human-Agent-Society/CORAL
Owner	Human-Agent-Society
Primary language	Python
License	Apache-2.0 — OSI-approved
Stars	783
Forks	101
Open issues	12
Latest release	v0.7.6 (2026-07-05)
Last updated	2026-07-07
Source	https://github.com/Human-Agent-Society/CORAL

What CORAL is

Multi-agent orchestration platform using isolated git worktrees per agent, symlinked shared state (`.coral/public/`), async grader evaluation, and heartbeat-driven prompts (reflect/consolidate/pivot). Supports LiteLLM gateway for model routing; requires Docker for certain graders (SWE-bench, terminal-bench) but must run on host, not in containers.

Quickstart

Get the CORAL source

Clone the repository and explore it locally.

terminalbash

1git clone https://github.com/Human-Agent-Society/CORAL.git2cd CORAL3# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Automated Algorithm Optimization

Leverage multiple agents exploring solution space in parallel (e.g., circle packing, kernel optimization). Grader provides fast feedback loop; agents consolidate insights across attempts.

Research & Proof-of-Concept Automation

Spawn agent teams on open-ended discovery tasks (math conjectures, ML hyperparameter tuning, bio sequence analysis). CORAL's shared knowledge model reduces duplicate work.

SWE-Bench & Coding Benchmarking

Orchestrate multiple coding agents against standardized benchmarks with containerized evaluation. Built-in Harbor integration handles test isolation and scoring.

Implementation considerations

Agent runtime (Claude Code, Codex, Cursor, Kiro, OpenCode) must be installed and authenticated separately; each has distinct config and model routing requirements.
Graders are pluggable Python entrypoints; custom grader requires packaging and wiring via `grader.entrypoint` (legacy auto-discovery removed as of v0.6.0+).
Docker required for SWE-bench and terminal-bench; CORAL itself must run on host (no DinD support). Multi-island runs (v0.6.0+) partition agents into scoped exploration zones.
Shared state (`.coral/public/`) is symlinked into all worktrees; agents are unprivileged in Docker (isolate_user mode) and cannot access grader secrets (`.coral/private/`).
Python 3.11+ required; install via `uv tool install` or manual setup. Development uses pytest, ruff for lint/format.

When to avoid it — and what to weigh

Deterministic, Audited Workflows Required — Agent behavior is non-deterministic. Regulatory/compliance contexts requiring full reproducibility and decision auditability will struggle with autonomous agent loops.
Real-Time Production Serving — CORAL is built for batch experimentation and research, not low-latency inference. Multi-agent evaluation loops are inherently slow.
Limited LLM Budget or Offline Environments — Each agent is an LLM client; parallel runs scale cost linearly. Requires live internet for Claude, Codex, Kiro APIs. No local-only mode documented.
Windows-First Development — Heavy reliance on bash, git worktrees, Docker, and POSIX primitives. Docker-in-Docker not supported. No explicit Windows compatibility stated.

License & commercial use

Apache License 2.0 (Apache-2.0). Permissive OSI-approved license allowing commercial use, modification, and distribution with liability and trademark disclaimers.

Apache-2.0 is a permissive open-source license. Commercial use, derivative products, and proprietary modifications are explicitly permitted, provided the license and copyright notice are retained. No copyleft requirement. Consult your legal team if you bundle with other licensed code or offer as a service.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

Signal	Assessment
Maintenance	Active
Documentation	Strong
License clarity	Clear
Deployment complexity	High
DEV.co fit	Good
Assessment confidence	High

Security considerations

Docker isolation (v0.6.0+) runs agents as unprivileged users; agents cannot read `.coral/private/` (grader venv, answer keys) even via bash. Grader and manager remain root in container. On-host isolation is opt-in (`agents.isolate_user`). No explicit mention of supply-chain security, dependency pinning, or vulnerability scanning. Multi-agent evaluation exposes agents to untrusted task specs and grader outputs; sanitization of grader I/O not detailed. Recommend threat modeling before production use.

Alternatives to consider

Anthropic's Claude Coder / Claude Code

Single-agent coding assistant; lacks multi-agent orchestration, persistent shared state, and autonomous loop capabilities. Use if you need a single interactive agent, not research-scale parallelism.

OpenAI Swarm / Bee Agent Framework

Lighter-weight multi-agent frameworks with simpler orchestration. CORAL is heavier and more opinionated (git worktrees, eval loops, grader daemons); Swarm is more modular for custom agent graphs.

LangChain / LlamaIndex Agents

General-purpose LLM agent libraries. CORAL is specialized for autonomous iterative refinement and benchmarking; LangChain is broader but less opinionated on eval-loop and multi-island logic.

Software development agency

Build on CORAL with DEV.co software developers

CORAL is open-source and production-ready. Start with the quick-install, scaffold a task, and launch agents. Review the docs and examples to understand grader authoring and multi-agent orchestration before scaling.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

SWE-AF agents-cli agents

Related on DEV.co

Explore the category and the services that help you build with it.

AI Coding Agents AI Application Development DevOps Custom Software Development Talk to DEV.co

CORAL FAQ

Can I run CORAL in Docker?

CORAL itself should not run inside Docker if you're using SWE-bench or terminal-bench (those require Harbor/Docker, and DinD is not supported). Run CORAL on the host. Individual agent workspaces can be containerized.

Do I need to pay for agent runtimes?

Yes, agent runtimes (Claude Code, Codex, Cursor, Kiro, OpenCode) are external services or paid products. CORAL orchestrates them but does not provide them. API costs accumulate with parallel agent runs.

Can I use local LLMs instead of cloud APIs?

Partially. LiteLLM gateway supports Ollama and other local models, but none of the built-in agent runtimes (Claude Code, Codex, Kiro) have documented local-only modes. Custom agent runtimes would require extension.

What happens if two agents edit the same file?

Each agent runs in a git worktree with its own working directory. Changes are committed and visible to other agents via symlinked `.coral/public/` state. File-level conflicts are resolved by git semantics and grader feedback; no explicit merge conflict resolution is documented.

Software developers & web developers for hire

From first prototype to production, DEV.co delivers software development services around tools like CORAL. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across ai coding agents and beyond.

Ready to Automate Agent Research?

Talk to DEV.co Schedule a Consultation