AI agent glossary: audit trail, guardrails, human-in-the-loop
Reelease’s own vocabulary and the terms the market uses for AI agents: what each one means and how Reelease handles it.
Reelease terms
The words you’ll see on the platform, and what each one means.
- Control layer
- Reelease’s category: the control layer between AI and your systems, through which agents act with rules, approvals and a record of every run.
- Also known as: AI agent governance
- Sealed Plan
- The plan that runs: the platform builds it from your use cases and your rules, starting from what the model proposes, seals it with a fingerprint and signs each step’s data. If anything changes, it doesn’t run.
- Also known as: signed plan
- Two-Key Approval
- The default rule for human approval: whoever starts a task can’t approve its changes. Someone else does, with the exact values in front of them. An admin can let the person who started it approve with a reason, and that goes into the run record.
- Also known as: four eyes, maker-checker, segregation of duties
- Pre-analysis
- The first thing Reelease does with a new source. In seconds it recognizes what kind of system it is, forecasts what you will be able to do with it and shows you what will be read and what is left out, before the analysis.
- Operable system map
- What Reelease knows about your system (code, screens, rules and use cases), reviewed by your team. Everything the agent does builds on it.
- Catalog
- The reviewable list of what your agent knows, under “Knowledge”: use cases, screens, connections, business rules and domains. What the analysis finds stays as a suggestion until your team reviews it, and only the use cases it publishes become executable.
- AI-generated documentation
- Complete documents (architecture, functional specification, user manual, executive summary) generated from the reviewed map, with sources.
- Executable use case
- A use case your team connects and publishes so the agent can run it on your systems, step by step and with its data. It’s the only thing an agent can run: the rest of the system is understood, but not touched.
- Documentation use case
- What the analysis understands and documents about your system (steps, data, screens and rules) without running it. It’s there to look up and to plan a replacement, and it only becomes executable if your team connects and publishes it.
- Connection
- What links an executable use case to the system that handles it today: a core transaction, an API or a database. When you replace that system, you change the connection and the use case stays the same.
- Client Bridge
- A component you install in your network so the agent can use your databases and procedures without exposing them to the internet.
- Task
- What an agent does for you: look up, prepare or change data in your systems, under your rules and the approvals you define. Each launch is a run, and every run leaves its run record.
- Work
- The chat mode that plans and runs tasks on real systems, within the agent’s rules.
- Ask
- The chat mode for querying the agent and co-writing documents with sources. It changes nothing in your systems.
- Operator
- Someone who works with their agents from the chat, using their own work accounts and with no access to the platform panel.
- Rules
- Each agent’s rules decide what it can do, which steps wait for a person and how far it can go. Only an admin can change them, every change is recorded and the agent cannot skip them.
- Also known as: guardrails
- Starts with no permissions
- A new agent can’t touch any system. An admin turns on, one by one, the capabilities it needs: your systems, email, documents, the browser or the operator’s computer.
- Also known as: least privilege
- Pre-run check
- What gets checked before every run: that the agent is active, that its rules allow every step and that the operator’s accounts and the place where the agent will work are ready.
- Human oversight
- In Reelease, you define it: which steps wait for a person (approve, choose or fill in a form), who decides and by when. By default, nothing changes data without someone approving it.
- Also known as: human-in-the-loop
- Emergency list
- The people who, with a justification, can approve a data change when the person who should approve it isn’t available.
- Run record
- The exportable record of every run: who started it, who approved each step, what was done in each system and under which rules.
- Also known as: audit trail
Market terms
What each one means, how Reelease handles it and what we call it.
Audit trail
An audit trail is a chronological record of who did what and when in a system, kept so any action can be reviewed later.
Reelease calls it the run record: an exportable record of every run, with each step, approval and action.
See a sample run recordGuardrails
Guardrails are the limits placed on an AI system so that it stays within what it is allowed to do.
In Reelease they are each agent’s rules: they set what it can do and what waits for a person, and the agent cannot skip them.
See each agent’s rulesHuman-in-the-loop
Human-in-the-loop means a person reviews or approves what an AI system does before the action goes ahead.
In Reelease, that means the approvals you define: which steps wait for a person, who approves and by when.
See the approvals you defineLeast privilege
Least privilege is the security principle of giving each user or system only the access it needs to do its job, and nothing more.
In Reelease, every agent starts with no permissions and an admin turns on the capabilities it needs, one by one.
See how every agent startsAI agent governance
AI agent governance sets what each AI agent may do, who oversees it and how its actions are recorded.
In Reelease it is the control layer: rules per agent, the approvals you define and a run record for every run.
See what you controlAgentic AI
Agentic AI is AI that plans and carries out tasks on real systems to reach a goal, rather than only answering questions.
In Reelease, agents work on the systems you already run and carry out only what your team publishes.
See what Reelease isEU data residency
EU data residency means data is stored and processed inside the European Union.
With the Reelease service, your systems’ information stays in the EU (Madrid and Frankfurt) and the AI models run inside the EU too.
See data and privacyOn-premises
On-premises software runs on the organization’s own servers and network, not in a vendor’s cloud.
Reelease runs on your premises too: platform, agents and AI model on your servers, and your data stays in your network.
See where your AI runsAir-gapped
An air-gapped system runs on a network with no connection to the internet or to any other outside network.
Reelease calls it isolated mode: no outbound connection, the AI model on your own hardware and signed updates.
See isolated mode
See it on your own system
We’ll show you Reelease on one of your cases: what it understands, the plan an agent would propose and what you’d sign off on.