Concurrency

AgentCompass has two concurrency layers: per-run task concurrency and process-global provider limits.

Per-Run Concurrency

agentcompass run \
  swebench_verified \
  mini_swe_agent \
  your-model \
  --env <env-provider> \
  --model-base-url "$MODEL_BASE_URL" \
  --model-api-key "$MODEL_API_KEY" \
  --task-concurrency 32

--task-concurrency controls how many benchmark tasks can execute in parallel for this run.

Provider Limits

Provider limits live in code defaults and can be overridden with runtime.provider_limits in private config files. They default to 128 per provider and cap concurrent sessions across the process, even if a run requests a higher task concurrency.

Limit	Controls
`task_concurrency`	Tasks in one run.
`runtime.provider_limits.<provider>`	Process-wide sessions for a provider.
Model endpoint rate limit	External API throughput and retry behavior.

Practical Workflow

Start with one sample_ids task.
Increase --task-concurrency gradually.
Watch model errors, provider startup time, and task timeout failures.
Keep provider limits below account or cluster quotas.

Get Started

Key Modules

Environments

Cookbooks

Per-Run Concurrency

Provider Limits

Practical Workflow

​Per-Run Concurrency

​Provider Limits

​Practical Workflow

​Related Pages

Per-Run Concurrency

Provider Limits

Practical Workflow

Related Pages