Beyond the Chatbox: LLM Coding Agents for Research

27 Glossary

Author

Affiliations

Agent: A model connected to tools through a harness loop.
Harness: Software that provides tools, instructions, permissions, context handling, and an interface.
Context: Instructions, history, tool results, and selected content sent to a model.
Skill: Reusable guidance for a specific task.
Hook: Code run at a harness event. It may transform data but is not an OS boundary.
Sandbox/container: An environment boundary that limits filesystem and process impact.
Prompt injection: Untrusted content trying to act as an instruction.
Data contract: Executable checks for schema, values, keys, and invariants.
Provenance: Records of inputs, transformations, versions, and outputs.
Memory: Maintained project records that help a later session resume.

Building effective agents — A readable source for the workflow, agent, tool, and loop vocabulary used throughout the glossary.
NIST’s Adversarial Machine Learning taxonomy — The deeper standardized reference for security terms such as prompt injection, attacker capability, and mitigation.