What Is a Hosted AI Agent? (And Why It Beats a Chatbot)
Most people's first experience with AI is a chatbot: you type a message, the model replies, the conversation ends. That interaction model is useful for answering questions — but it's fundamentally broken for running a business.
A hosted AI agent is a different animal entirely. It doesn't wait for you to say something. It runs continuously in the background, watches for events, takes actions, persists memory across sessions, and loops in humans when decisions matter. It's the difference between a very smart FAQ bot and an autonomous employee that actually gets things done.
This post explains exactly what a hosted AI agent is, how it differs from a chatbot, and why that distinction determines whether AI can automate real workflows for you.
The Core Problem With Chatbots
Chatbots are stateless and passive. Every conversation starts from zero. The bot has no memory of what you discussed last Tuesday. It doesn't know what happened in your system while you weren't chatting. It can't do anything unless you prompt it first.
This works fine for customer support FAQs or a simple Q&A interface. But most meaningful business automation doesn't start with a human typing a question. It starts with an event:
- A bug alert fires in Sentry
- A new lead appears in your CRM
- A scheduled report is due
- A customer leaves a negative review
A chatbot misses all of these. It's not listening.
What Makes an Agent "Hosted"
A hosted AI agent runs in a persistent infrastructure environment — not a serverless function that spins down between requests, but a real long-running process with:
- Always-on execution — the agent is running 24/7, listening for triggers
- Persistent memory — state is preserved across sessions (semantic memory, structured records, key-value store)
- Event-driven activation — fires on webhooks, database change events (CDC), schedules, or messages from other agents
- Tool access — calls APIs, reads databases, sends emails, creates tickets, writes code
- Human-in-the-loop gates — pauses execution to get human approval when needed, then resumes automatically
The "hosted" part matters because you don't manage the infrastructure. The platform handles uptime, scaling, restart policies, and tool connectivity. You define what the agent does; the platform keeps it running.
How SchemaBounce Runs Hosted Agents
On SchemaBounce, every agent runs as a persistent process in the Agent Data Layer (ADL) — a purpose-built memory and data system for AI agents. Here's what that looks like in practice:
Event fires (CDC, webhook, schedule)
↓
Agent wakes, reads its memory (namespaces: working, durable, ephemeral)
↓
Queries entity records (contacts, deals, incidents, etc.)
↓
Takes action (calls tools, updates records, messages other agents)
↓
[Optional] Escalates to human for approval
↓
Resumes execution after approval, logs outcome
This isn't a prompt-response loop. It's a stateful process with a full data layer behind it.
Memory That Actually Persists
SchemaBounce agents have three memory tiers:
| Tier | Lifecycle | Use Case |
|---|---|---|
ephemeral | 1 day | Per-run scratch data |
working | 7 days | Active task context |
durable | Indefinite | Long-term learned facts |
Memory is stored in namespaces and is searchable semantically — the agent doesn't just remember raw text, it can retrieve relevant facts by meaning, not just exact key lookup. This is the difference between an agent that "knows" your preferences and one that forgets every session.
Triggers, Not Prompts
Instead of waiting for a user message, a SchemaBounce agent can trigger on:
- CDC events — database row changes (e.g., new deal created in your CRM)
- Webhook payloads — Sentry alert, Stripe payment, GitHub PR
- Cron schedules — daily at 9am, every hour, every Monday
- Agent messages — another agent sends a
requestoralert
You can mix these. An incident-response agent might wake on a Sentry webhook, query related records, attempt an automated fix, and only escalate to a human if confidence is below a threshold.
Agent vs Chatbot: Side-by-Side
| Capability | Chatbot | Hosted AI Agent |
|---|---|---|
| Runs without user prompt | ❌ | ✅ |
| Persistent memory across sessions | ❌ | ✅ |
| Event-triggered execution | ❌ | ✅ |
| Multi-step tool use | Limited | ✅ |
| Human approval gates | ❌ | ✅ |
| Coordinates with other agents | ❌ | ✅ |
| Connected to your data layer | ❌ | ✅ |
| Runs 24/7 on managed infra | ❌ | ✅ |
The gap isn't incremental. A chatbot is a UI component. A hosted agent is an automated system participant.
Real-World Example: Incident Response
Here's what a chatbot-based incident flow looks like:
- Sentry fires an alert → engineer sees it 20 minutes later
- Engineer asks ChatGPT to help debug → pastes error log manually
- Engineer files a Jira ticket, notifies Slack, checks runbooks
- Engineer deploys a fix after 2+ hours
Here's the same flow with a hosted agent:
- Sentry fires a webhook → agent wakes in seconds
- Agent queries related records, checks recent deploys, reads runbook from memory
- Agent files Jira ticket, posts to Slack with diagnosis, attempts automated rollback if confidence > 0.85
- Agent escalates to on-call engineer only if automated remediation fails
- Full incident log written to ADL for post-mortem
Same problem. Different order of magnitude in speed and engineer load.
The 90+ MCP Servers Factor
SchemaBounce ships with 90+ Model Context Protocol (MCP) servers pre-integrated — Sentry, Jira, Slack, GitHub, Stripe, Salesforce, Linear, and many more. Agents connect to your existing tools without custom integration code.
This is what "hosted" unlocks at scale: not just an agent that can reason, but one that's already wired into the tools your team uses every day.
When to Use a Chatbot vs a Hosted Agent
Use a chatbot when:
- You need a simple Q&A interface
- Users drive all interactions
- No background processing needed
- Stateless responses are fine
Use a hosted agent when:
- You need automation that runs on its own schedule
- State must persist between interactions
- You're connecting to external systems and APIs
- Multiple systems need to coordinate
- Human approval is required at specific decision points
Getting Started
SchemaBounce's bot marketplace has 64 pre-built agents ready to deploy — incident response, SEO analysis, content creation, lead enrichment, and more. Each runs on the hosted ADL infrastructure from day one.
No infrastructure setup. No agent loop to build. The agent runs when you deploy it.
If you've been evaluating whether AI can actually automate meaningful work for your team — the answer is yes, but only with a hosted agent, not a chatbot.
Want to go deeper? Read: What Is the Agent Data Layer (ADL)? and How to Build a Multi-Agent Team With Human Approval Steps.