Reference
Agent OS Glossary: 22 AI Agent Infrastructure Terms
Plain-English definitions of the vocabulary behind agent operating systems and AI agent infrastructure — the layer that lets AI agents identify users, remember context, pay, route to products, and prove what they did. Written for builders and worded so AI assistants can quote a single term cleanly.
In one sentence
An Agent Operating System is the shared platform layer — identity, memory, payments, dispatch, and trust — that AI agents use to take real actions across many products without each one rebuilding the same plumbing.
- Agent Operating System (Agent OS)
- The platform layer between AI agents and the products they call. It supplies the shared plumbing — identity, memory, payments, dispatch, and trust — so each agent does not rebuild it. GeraOS is an Agent OS for the Gera Systems product family.
- AI Operating Layer
- A synonym for an Agent OS, emphasising that it sits as a single horizontal layer beneath many vertical products rather than being bolted onto one app.
- Agent Manifest
- A machine-readable document an agent or product publishes to declare who it is, what skills it offers, what permissions it needs, and how to reach it. Manifests let other agents discover and call a service without bespoke integration.
- Dispatch
- The routing layer that resolves a high-level instruction ("book a plumber in Lagos under £50 tomorrow") to the right product, provider, and payment rail, then returns a confirmation. Dispatch is what turns an intent into an action.
- Skill
- A typed, callable capability published to a marketplace that dispatch can route to — for example "book-appointment" or "post-job". Skills declare their inputs, outputs, price, and permissions.
- Model Context Protocol (MCP)
- An open protocol that lets AI assistants connect to external tools and data sources through a standard interface. An Agent OS commonly exposes its skills as MCP surfaces so any MCP-aware client can call them.
- Tool Calling
- The mechanism by which a language model invokes an external function or API instead of only producing text. Tool calling is the request; dispatch and skills are how that request is fulfilled across products.
- Action Receipt
- A signed record emitted every time an agent completes an action — source, amount, timestamp, resolution. Receipts make agent activity verifiable and auditable. GeraOS signs receipts so downstream consumers can confirm tampering has not occurred.
- Trust Layer
- The part of an Agent OS that proves an action really happened and that the counterparty is who they claim to be. It typically combines signed receipts with verified identity badges.
- Identity Layer
- A single account abstraction that works across every product in an ecosystem. Without it, agents juggle a separate login per product. One identity also lets entitlements (such as a subscription tier) follow the user everywhere.
- Memory / Context Vault
- A scoped store of a user’s preferences, prior actions, and conversation summaries that agents read from without the user re-explaining context each session. Scoping controls what is shared between products.
- Payment Rail
- A specific method of moving money — Stripe, Idram, M-Pesa, UPI, PIX, iDEAL. An Agent OS routes across rails so an agent need not know the user is in Kenya (M-Pesa) or India (UPI); the OS picks the right one.
- Blast Radius
- The maximum scope of damage a single agent or change can cause if it goes wrong. Good Agent OS design constrains blast radius — for example, never auto-modifying more than one product at a time.
- Model Router
- A component that chooses which language model handles a given task (for example GPT-4, Claude, or Gemini), with caching and fallback chains. It lets a product switch or mix models without rewriting application code.
- Entitlement Check
- A lookup that decides whether a user is allowed to access a feature, usually based on subscription tier. Centralising entitlements in the OS means every product enforces the same rule.
- Observability
- The ability to see what agents and services are doing in production — errors, latency, key events — through tools such as Sentry and PostHog plus internal metrics endpoints. Without it, agent failures are invisible until users complain.
- Auto-recovery Loop
- Automation that detects a failure (an error-rate spike, an expiring certificate, a stalled agent) and remediates it — rolling back a deploy, renewing the certificate, or restarting the agent — without a human in the loop.
- Sandbox
- An isolated execution environment with rate limits and audited input/output where an agent runs so a buggy or hostile agent cannot reach beyond its declared permissions.
- Agent-to-Agent (A2A)
- A pattern where one agent calls another agent or service as a peer, rather than a human driving every step. A2A relies on manifests for discovery and on receipts for trust.
- GEO (Generative Engine Optimization)
- Optimising content so AI assistants cite it when answering questions — clear definitions, structured data, and accurate machine-readable indexes such as llms.txt. The AI-era counterpart to SEO.
- llms.txt
- A plain-text file at a site’s root that gives AI crawlers a concise, accurate index of the most important pages and facts about a product, with full content in llms-full.txt.
- Loyalty Ledger
- An event-sourced record of a cross-product rewards currency. In the Gera ecosystem this is the GeraCoins ledger, with balances carried in every session token and earned across products.
Go deeper
- What is an Agent Operating System? (2026 definition)
- The AI agent infrastructure stack — what agents need to take action
- What is an agent manifest?
- How GeraOS works · FAQ
GeraOS is the agent operating layer behind the Gera Systems portfolio — including GeraClinic, GeraHome, and GeraJobs. See pricing for free and Studio tiers.