| name | aire-agent-workflow |
| description | Orchestrates work on the University of Leeds AIRE HPC cluster using the aire-agent toolkit. Use when starting AIRE sessions, submitting Slurm jobs, choosing between raw Slurm and aire-agent tools, wiring MCP for Claude Code, or following generate-validate-submit-debug loops on AIRE. |
AIRE Agent Workflow
Operating mode
Act as an AIRE cluster assistant. The install root is ${AIRE_AGENT:-$HOME/.aire-agent} (call it AIRE below).
Start substantial tasks with a short note: task type, partition/GPU needs, and whether the user is on a login node or local machine.
Core rules (never violate)
- Max 3 GPUs per node (
--gres=gpu:N, N≤3); use multi-node for more.
--partition=gpu and --gres=gpu:N together on GPU jobs.
--time required on every job.
- Default Slurm allocation is 1 CPU, 1 GB — request more explicitly.
- No training on login nodes — submit jobs or use
srun.
$TMP_SHARED is wiped when the job ends — copy artifacts to $SCRATCH.
- Conda envs live in
$HOME; datasets/checkpoints on $SCRATCH.
Read $AIRE/knowledge/ and $AIRE/AGENTS.md before guessing policy.
Slurm — run directly
| Task | Command |
|---|
| Submit | sbatch script.sh |
| Queue | squeue --me |
| Cancel | scancel <job_id> |
| Efficiency | seff <job_id> |
| Quota | quota -s |
| GPU nodes | sinfo -p gpu |
Do not wrap these in custom tools unless the user asks for CLI helpers (aire-agent queue, etc.).
aire-agent CLI — value-add only
export PATH="$AIRE/bin:$PATH"
aire-agent generate --gpu 1 --time 4h --framework pytorch
aire-agent validate my_job.sh
aire-agent search "GPU memory"
aire-agent log --name run1 --metrics '{"loss":0.3}' --params '{"lr":0.001}'
aire-agent doctor
MCP server (Claude Code and compatible clients)
Register once:
claude mcp add aire "$AIRE/mcp/server.py"
Registered MCP tools (see $AIRE/mcp/server.py):
| Tool | Purpose |
|---|
generate_script | SBATCH with AIRE constraints |
validate_script | Pre-submit checks |
search_docs | Knowledge base search |
list_modules | Module list |
system_info | Hardware summary |
log_experiment | JSONL experiment log |
query_experiments | Query past runs |
sync_docs | Refresh upstream docs |
Job queue/submit tools are intentionally not in MCP — use Slurm directly.
Standard workflow
- Environment — skill
aire-conda-environments; optional aire-github-installs.
- Generate —
aire-agent generate or MCP generate_script.
- Validate —
aire-agent validate or MCP validate_script.
- Submit —
sbatch script.sh.
- Monitor —
squeue --me; tail logs/.
- After completion —
seff <job_id>; skill aire-ddp-debugging if distributed.
- Log run —
aire-agent log or MCP log_experiment.
Skill routing
| User need | Load skill |
|---|
| Conda / env YAML | aire-conda-environments |
| GitHub pip / editable install | aire-github-installs |
| Multi-GPU training | aire-l40s-distributed-training |
| DDP hang / NCCL | aire-ddp-debugging |
| Repo layout / Hydra | aire-research-software-engineering |
| Medical CV science | medical-cv-research-engineer (domain, not AIRE ops) |
References
$AIRE/CLAUDE.md — Claude Code session behavior
$AIRE/AGENTS.md — generic agent rules
$AIRE/README.md — install and examples