Glossary
Plain-language definitions for terms that come up across ossbeat posts. Linked wherever they first appear in an article.
A2A protocol (Agent2Agent)
A communication standard that lets one AI agent call another AI agent directly, rather than routing everything through a human or a single orchestrator. An agent can publish itself as callable, and other agents can discover and invoke it.
Agent harness
The scaffolding around a coding agent (like Claude Code or Codex) that adds skills, memory, safety rules, and conventions on top of the base agent, without changing the agent itself. Think of it as a configuration layer that makes an agent more capable and consistent across projects.
Checkpoint (agent workflows)
A saved snapshot of an agent's state at a specific point in a task, letting the agent pause and later resume exactly where it left off — used for human approval steps, error recovery, or long-running jobs.
Human-in-the-loop (HITL)
A design pattern where an agent pauses before taking a consequential action (like sending an email or writing to a database) and waits for a person to approve, reject, or edit it, instead of acting fully autonomously.
LLM gateway
A layer that sits between an application and one or more language models, routing requests to whichever model or provider is configured — local or cloud — without the application needing to know the difference.
MCP (Model Context Protocol)
An open standard (created by Anthropic) for connecting AI models to external tools, data sources, and services through a consistent interface, so a tool built for one MCP-compatible agent can generally work with others.
Prefill
The initial processing step where a model reads and encodes a prompt before it starts generating a response. Efficient prefill handling matters most when a generation is cancelled and resumed, or when reusing context across requests.
Prompt injection
An attack where instructions hidden in content an AI system processes (a webpage, a document, an email) attempt to override the system's actual instructions. Tools that read untrusted content need specific defenses against it.
Structured output
A mode where a model's response is constrained to a specific format (usually JSON matching a schema) instead of free-form text, making it reliably parseable by other code.