coral-quickstart

CORAL quickstart

CORAL is infrastructure for autonomous coding agents: you give it a codebase (seed/) and a grader (turns a commit into a number), and it spawns agents in isolated git worktrees that edit code, submit commits, and get scored on a shared leaderboard — looping to push the score up. The agents are the optimizer; your grader defines "better".

When to reach for CORAL

Good fit:

• You can express success as a number — accuracy, runtime ratio, pass rate, or a rubric-judge score for open-ended work.
• The work is iterative search: many attempts at one well-scoped problem (kernel/algorithm optimization, benchmark solving, prompt/program tuning, "make this function faster").
• You want parallel agents exploring independently and sharing what works.

Not a fit:

• One-shot tasks with no measurable objective.
• Work that needs a human judging every attempt (use a rubric-judge grader if a model can score it).

How a run is shaped

you provide:   seed/ (starter code)  +  a grader (commit → number)
coral spawns:  N agents, each in its own git worktree
each agent:    edit code → `coral eval` → grader scores it → read leaderboard → repeat
shared state:  attempts, notes, and skills are visible across agents in real time

Two things you build (seed/ + grader) and one thing you tune (how many agents, which model). Worktrees, scoring daemon, shared state, and restarts are handled for you.

Get running — four steps

1. Install the CLI

curl -fsSL https://raw.githubusercontent.com/Human-Agent-Society/CORAL/main/install.sh | sh
# or, if you have uv:  uv tool install coral
coral --help      # verify

The installer grabs the latest coral release by default — that's what you want. (Only pin a specific release with CORAL_VERSION=<tag> before the curl if you have a reason to.)

2. Register a runtime (coral setup)

coral shells out to a coding-agent CLI (Claude Code, Codex, Cursor, Kiro, OpenCode) — each installed and authenticated separately. Tell coral which to use:

coral setup            # scans PATH, wizard to create named runtime bindings
coral agents doctor    # validates them, incl. a live auth ping

If doctor's live ping fails (expired auth, model typo, "runtime not found" at start) → the setting-up-coral skill has the full troubleshooting matrix.

3. Build a task — use a .coral_workspace/

When the user wants CORAL to optimize code they already have, keep every bit of CORAL scaffolding (task config, seed, grader, results) inside a .coral_workspace/ directory at the root of their project. This keeps coral out of their actual source tree and is trivially gitignored.

Act — never answer "optimize this" with a menu. A request like "use coral to optimize this" (in any language) is a build instruction, not a question. The failure mode to avoid is a process menu like "1. point me to a task 2. create one 3. optimize outside coral" — do not produce that. A repo with no task.yaml isn't ambiguous; it just means you build the task from the current repo. So:

1. Dig for what's already measurable. Research/framework repos (like SAGA) almost always ship an eval/benchmark script, a test suite, or a headline metric in the README/paper. Find it — that's your optimization target and metric.
2. If no single number is obvious, construct one by wrapping the repo's existing evaluation. Don't give up and ask just because there's no CORAL scaffold.
3. Scaffold the most plausible target and start building — a .coral_workspace/ + draft grader is cheap and reversible. State your assumption in one line ("Optimizing SAGA's from
4. Last resort only: if you've actually read the repo and it exposes nothing scorable, propose 2-3 concrete targets you found (each with a metric), pick the most likely, and scaffold it — still not a process menu.

Stop before coral start (a paid run); everything up to a validated task is autonomous. On Claude Code, delegate this whole grind to the coral-task-author subagent.

The mechanical boilerplate (gitignore + coral init + copy the code into seed/) is bundled as a script — run it from the project root:

"${CLAUDE_PLUGIN_ROOT}/skills/coral-quickstart/scripts/new-coral-workspace.sh" optimize path/to/their_module.py

It scaffolds .coral_workspace/optimize/ and copies the file into seed/solution.py, leaving only the grader for you to write. (Equivalent by hand:)

# from the user's project root
echo ".coral_workspace/" >> .gitignore
mkdir -p .coral_workspace && cd .coral_workspace
coral init optimize && cd optimize
cp ../../path/to/their_module.py seed/solution.py   # the code to optimize

Then make the task fit the user's goal — two edits:

• task.yaml → set task.description to what the agents should optimize and the program file's contract (e.g. "solution.py must define run() and stay correct; we score speedup").
• the grader → score the user's actual metric (speedup vs baseline, accuracy on a held-out set, pass rate, …). This is the heart of it → the creating-a-coral-task skill walks through grader patterns and the TaskGrader API.

4. Validate, then launch

coral validate .                 # confirms the grader scores the seed — the one checkpoint that matters
coral start -c task.yaml         # launch agents (results stay under .coral_workspace/)
coral status                     # watch the leaderboard

If coral validate succeeds, the grader can score the seed; most "agents are stuck" reports trace to a grader that crashes here. Driving the run from here — monitoring, steering, stopping — is the running-coral-experiments skill.

The workflows (where to go next)

You want to...	Skill	Commands
Set up runtimes (one-time)	setting-up-coral	coral setup, coral agents doctor
Author a task / write a grader	creating-a-coral-task	coral init, coral validate
Run / manage experiments	running-coral-experiments	coral start / status / log / show / resume / stop

The eval loop inside a run (coral eval -m "..." → score → iterate) is driven by the agents themselves — they read it from the CORAL.md CORAL generates, so you never run it by hand. Docs: https://docs.coralxyz.com/