ML Intern

ML Intern is a command-line agent for machine-learning work against a model hub.¹ Its tools run on the local filesystem or in a remote sandbox, chosen by trust and scope.¹

Overview#

The loop checks for a doom loop and can ask for approval before a sensitive operation.¹ Compaction is an explicit event, not a silent death of context.¹

Mechanism#

1. Local filesystem tools or sandbox tools are chosen by trust and scope.¹
2. The loop iterates on completion or tool calls and checks for a doom loop.¹
3. Approval is requested before a sensitive operation.¹
4. Compaction is an explicit operation that keeps a trace.¹
5. Extra tool servers are an option, not the default.¹

Applications#

The agent is relevant when a research agent needs a local-versus-sandbox tool choice and a stall check.

Limitations#

A fat remote-tool layer is not the default path.¹

See also#

• Harness runtime – runtime internals
• Verification and stop conditions – doom-loop checks and approvals
• Autoresearch loop – machine-learning iteration loops at a different layer
• DataClaw – related data-science glue

References#

Footnotes#

1. https://github.com/huggingface/ml-intern ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰