harbor run flags while supporting multiple agents, datasets, and tasks in one reusable JSON or YAML file.
Example:
config.yaml
Creating a config
You can useharbor run.
To generate a config with all available fields:
--print-config to inspect the resolved config without running the job:
harbor job schema to print the config JSON Schema:
--dry-run to check the config, agent kwargs, credentials, and task sources without downloading tasks or running the job:
--launch to use Harbor Hub schemas and validation.
All top-level fields are optional. Harbor applies the defaults shown below. Fields marked as required are required only when their containing object is present.
Job
Job name. When omitted, Harbor uses the current time in
YYYY-MM-DD__HH-MM-SS format.Multiplier applied to task timeouts unless a phase-specific multiplier overrides it.
Maximum number of concurrent trials. Must be at least 1. Per-agent concurrency limits cannot exceed this value.
Override
should_retry, should_grade, or has_valid_reward by exception type. Omitted fields keep their defaults. See Error classification.Optional simulated-user agent and bridge applied to every trial.
Environment paths collected after each trial. A string is shorthand for an artifact with only
source set.Inline text appended to each task instruction after
extra_instruction_paths.Source jobs used for regrading. When set, Harbor derives one new trial from each matching source trial instead of expanding tasks, agents, and attempts normally.
Retry
Exception class names eligible for retries.
null includes every exception not excluded below.Exception class names that are never retried. Exclusions take precedence over inclusions. The defaults are the non-retryable types in Error classification.
Agents
Per-agent limit on concurrent
agent.run() phases. Must be at least 1 and cannot exceed n_concurrent_trials.Shared concurrency-pool name. Agents in the same group must set the same
n_concurrent value.Resume the agent’s native session between steps of a multi-step task. Requires agent resume support.
Load a native
.jsonl or ATIF .json trajectory before the first step. Requires the corresponding agent capability.Hostnames, IP addresses, or CIDR ranges added to the allowlist during
agent.run() only.Glob patterns excluded from downloaded agent logs after applying
include_logs.MCP servers
MCP transport. The legacy value
http is normalized to streamable-http.HTTP request headers for
sse and streamable-http servers. Supported by claude-code and codex. Values support embedded ${VAR} and ${VAR:-default} references resolved before the agent is set up. Credential headers must use ${VAR} references. Other agents reject configurations with headers. See MCPs.Simulated user
user_agent supports every field in agents[], plus the fields below. The bridge field is required when user_agent is configured.
See Simulate a user.
Environment
Pre-integrated environment provider:
docker, podman, daytona, e2b, modal, runloop, langsmith, ec2, gke, ack, openshift, novita, apple-container, singularity, islo, tensorlake, cwsandbox, use-computer, cua-cloud, blaxel, opensandbox, beam, skypilot, hf-sandbox, hyperbrowser, vercel, sail, superserve, or sbx. See Pre-integrated sandboxes.How the provider enforces the task’s CPU value.
How the provider enforces the task’s memory value.
Replace the task’s CPU value at runtime. Fractional values require provider support.
Deprecated. This field is accepted but has no effect and is excluded when serializing configs.
Docker Compose long-syntax volume mounts for the agent environment only.
Hostnames, IP addresses, or CIDR ranges added to the environment network baseline.
TPU override
Mounts
Set to
false to prevent Docker Compose from creating a missing bind source path.Verifier
Glob patterns selecting verifier log files to download. The reward file is always downloaded.
Hostnames, IP addresses, or CIDR ranges added to the allowlist during
verify(), and from startup in a separate verifier environment.Datasets
Local dataset directory. With
repo, this selects a repo-relative implicit dataset directory.Harbor Hub dataset in
org/name format, or a bare dataset name for a custom or Git-repository registry.Version selected from a JSON registry or a named Git-repository registry.
Path to a custom
registry.json. With repo, this path is relative to the repository.Maximum number of tasks after inclusion and exclusion filters are applied.
repo, set either path or name, but not both. With repo, path selects an implicit dataset while name selects a named registry dataset. version and ref cannot both be set.
See Datasets, Registries, and Git repos.
Tasks
Local task directory, or the path within a Git repository when
git_url is set.path or name, but not both. A Git task uses path with git_url; a Harbor Hub task uses name with an optional ref.
See Tasks → Overview.
Metrics
Metric implementation used to aggregate task rewards.
Keyword arguments passed to the metric implementation.
uv-script requires script_path.Artifacts
Optional path under the trial’s artifact directory. Harbor derives it from
source when omitted.Docker Compose service to collect from.
null and main target the agent container.Regrade sources
Source job UUID. Required for a Hub source and optional for a local source.
Source job directory. Required for a local source and invalid for a Hub source.
install_only. See Regrade.
