Hub / Glossary
Lexicon
AI operations glossary
Yes. You can read any multi-agent ops page faster when each industry term maps to one plain meaning, and this glossary pairs the words operators search with the names we use inside the open-source 0xRay agent kit. Start with the industry column; our names are secondary, so you can bring your own.
Why map industry terms to our names?
Because you search in industry words, and our stack uses short internal names. Every guide on this site leads with the industry term, such as agent orchestration, observability, or human-in-the-loop, and mentions our name only as a gloss. This table is where the two meet, so a term you see in a guide or in the suit's docs points back to a concept you already know. For the bigger picture, read what AI fleet ops means.
Which industry terms map to which of our names?
| Industry term | Plain meaning | Our name (secondary) | Peer reference |
|---|---|---|---|
| Multi-agent system / agent fleet | Several agents run as one job | fleet, or house | Gartner on agentic AI |
| AI agent role | A named agent with one job | seat (for example coordinator, builder, reviewer, writer) | Microsoft CoE roles |
| Orchestrator / coordinator agent | The agent that assigns and routes work | CoS (chief-of-staff) seat | AgenRACI |
| Control plane / work board | One surface for status, owners, and handoffs | board, called Bifrost when live (not the unrelated LLM gateway of the same name) | Microsoft Entra AI control plane |
| Task ticket / work item spec | What to do, how to know it's done, who's next | card, plus its packet (goal, constraints, acceptance, gates, evidence, next owner, escalation, RACI) | — |
| RACI / one accountable owner | Exactly one person answers for the outcome | packet raci field; board flags raci_missing | Microsoft CoE roles |
| Human-in-the-loop / approval gate | A risky step waits for a named person | handoff kinds: ready, reviewer fail, live pass, merge, CI fail | LangChain HITL, OpenAI Agents SDK |
| Escalation / timeout ladder | Unanswered requests move to the next person | nudges, three levels 15 minutes apart by default, coordinator woken at the third | AgenRACI spec |
| Reviewer agent / eval gate | An independent pass or fail before merge | critic (PASS or FAIL) | AgenRACI |
| Release gate / CI quality gate | Work can't merge until checks pass | foundry gate (npx @0xray/foundry gate, then --verify-only) | — |
| Ship-readiness review | How strict the review is, by risk | mill gate at Light, Normal, or Strict | — |
| Agent observability / tracing | A record of every model and tool call | board activity feed and inbox lines, alongside OpenTelemetry traces | OTel GenAI conventions |
| Session state / checkpoint | Where an agent resumes after a reset | station card | LangGraph checkpointer (docs) |
| Agent identity / credentials | Each agent acts under its own key | per-seat key; a mismatched seat is refused | Microsoft Entra Agent ID |
| Budget cap / kill switch | Stop new work at a spend line | button up (weekly burn hold line) | Runaway loops |
| Agent harness / guardrail kit | The rules and tools each agent wears | suit (the kit, not the whole platform) | — |
| Sprint / milestone | A batch of work shipped together | wave | — |
| Engagement stages | First look, first working team, ongoing help | Scout, Stand-up, Retain | — |
How should I use these terms in my own team?
Use whichever names your team already says out loud, and keep a single glossary so each name means one thing. The concept is what matters: one accountable owner per action, a gate before risky steps, a trace per run, and a cap that stops new work. Our guides to map the agentic control plane and release gates for agents put several of these terms to work.
A note on two collisions: "AgentOps" is both a generic term and a product name, so we use it sparingly, and "Bifrost" also names an unrelated LLM gateway, so we say "our board" in public copy first.
How do you request a consult?
Email consult@0xray.dev.
What else do operators ask?
What is AI fleet ops?
It's the practice of running several AI agents as one accountable operation, with named owners, review gates, and a live work queue.
Is a control plane the same as an orchestration framework?
No. A framework runs agents; a control plane sets who owns the work, what is allowed, and where the record lives.
What does RACI mean for AI agents?
It assigns who does, who answers for, who advises on, and who hears about each action, with exactly one accountable human.
Do I have to use your names to work with you?
No. We map to your team's words; our names are shorthand inside the open-source suit.