# What is Event-Driven Agent?

Also called triggered agent, reactive agent.

An event-driven agent starts in response to something happening, such as an incoming message, a webhook, a record changing, or a threshold being crossed. Its trigger carries the initial context, and its value is responding close to the moment the event occurs. Volume is unpredictable, so limits and deduplication matter more than in scheduled work.

The pattern has three parts: a source that emits events, a filter that decides which events are worth a run, and the agent that handles the ones that pass. The filter is where most of the engineering sits, because starting a run for every event is expensive and noisy, while filtering too aggressively means the interesting cases never reach the agent.

Event systems have characteristic failures. The same event can be delivered twice, so handlers must be idempotent. A burst can trigger hundreds of concurrent runs, so concurrency caps and queues are required. And an agent whose own actions produce events that trigger it again creates a feedback loop, which is why triggers normally exclude changes the agent itself made.

Event payloads are untrusted input. An incoming message, a form submission, or an updated record can contain text written to influence the agent that will read it, and that text arrives with all the authority of a legitimate trigger. Payloads should be treated as data to be handled, never as instructions to be followed, with permissions enforced outside the prompt.

Compared with a scheduled agent, an event-driven one responds sooner and does less useless work, at the cost of unpredictable load. Compared with an interactive agent, it has no one present to ask, so the same unattended constraints apply. Many production systems combine all three, with events for responsiveness and a schedule that sweeps up whatever was missed.

## Key points

- Triggered by an occurrence, with context arriving in the payload.
- Filtering decides which events are worth a run at all.
- Duplicate delivery and bursts require idempotency and caps.
- Payloads are untrusted data, never instructions to follow.

## In practice

A form on a website is submitted. The event fires, and a filter checks that the submission is not a duplicate and that the company field is not empty. The agent then looks up the company, decides whether it matches the target profile, writes a short qualification note, and adds it to the record. Elapsed time from submission to note is under a minute.

## Related terms

- [Scheduled Agent](/en/glossary/scheduled-agent)
- [Autonomous Agent](/en/glossary/autonomous-agent)
- [Agent Orchestration](/en/glossary/agent-orchestration)
- [Agentic Workflow](/en/glossary/agentic-workflow)
- [Tool Use](/en/glossary/tool-use)

[Back to the AI Glossary](/en/glossary)