Plain definitions
Glossary
The vocabulary, defined without marketing.
Half the confusion in agentic AI is that the same word means three things depending on who is selling. These are the definitions we use internally.
Missing a term? Tell us and we will add it.
- Agent
- A system that decides its own next step and then takes it. The deciding is what separates it from a chatbot.
- Topic
- A scoped job an agent knows how to do, with its own instructions and actions. Overlapping topics are the most common design error we see.
- Action
- A specific thing the agent can execute — a Flow, an Apex method, a prompt template, an API call. Keep them small.
- Atlas Reasoning Engine
- The component that interprets the request, selects a topic, and plans the sequence of actions.
- RAG
- Retrieval-augmented generation. The agent looks something up before answering instead of relying on the model's memory.
- Agentic RAG
- Retrieval that plans: the agent decides what to search for, judges whether the result answers the question, and searches again if not.
- Grounding
- Tying an answer to a specific retrieved source. Ungrounded answers are the reason people distrust agents.
- Trust Layer
- Salesforce's guardrail stack — masking, zero retention, toxicity screening, grounding checks, audit trail.
- MCP
- Model Context Protocol. An open standard for exposing a system's capabilities as tools an agent can call safely.
- Agent2Agent (A2A)
- A protocol for agents built on different platforms to delegate work to each other and return structured results.
- Orchestration
- Deciding which agent owns which decision, what context passes between them, and who wins when they disagree.
- Guardrail
- A hard constraint the agent cannot reason its way past. If it lives only in the prompt, it is not a guardrail.
- Deflection
- A contact resolved without a human. Easy to measure, easy to game — check the repeat-contact rate alongside it.
- Containment
- The customer stayed with the agent to the end and did not need to escalate. A better metric than deflection.
- Prompt injection
- Hostile instructions hidden in content the agent retrieves. Treat all retrieved content as data, never as instruction.
- Eval
- A test case with a known-good outcome, run automatically against the agent. Without these, you are shipping on vibes.
Also in resources
Written by the people who build
Blog / Insights
Field notes from live deployments, including the ones that went badly.
Read Long-form, practicalGuides & Playbooks
The documents we use internally, lightly edited so they make sense outside the building.
Read Customer storiesCase Studies
Named where we are allowed to name them, and specific where we are not.
ReadStart with the assessment.
Three to four weeks, a fixed fee, and an answer we are willing to say out loud — including if the answer is no.