I've been prototyping a small internal tool to monitor Twitter (or X, whatever) for mentions of our product and specific keywords related to outages. The goal was simple: listen for tweets, filter them, and send a formatted alert to a Slack channel. I figured using Relevance AI's workflow engine would be a clean way to orchestrate this without managing a bunch of custom code.
I built the agent with a simple three-step workflow:
1. A trigger using their Twitter source connector, set to poll every 10 minutes.
2. A filter step to check tweet text against a list of terms.
3. A Slack step to post the formatted message.
Here's the basic YAML structure I used for the agent definition:
```yaml
agent:
name: twitter-outage-monitor
trigger:
type: source.twitter
config:
search_query: "(ourproductname OR #ourhashtag) (down OR error OR broken)"
poll_interval: 600
steps:
- id: filter_relevant
type: filter.text
config:
field: tweet.text
operator: contains_any
values: ["outage", "down", "failed", "bug"]
- id: post_to_slack
type: action.webhook
config:
url: ${SLACK_WEBHOOK_URL}
body: |
New potential issue tweet: {{tweet.text}}
From: @{{tweet.username}}
Link: https://twitter.com/user/status/{{tweet.id}}
```
It works perfectly. But then I got the bill for the first month. The cost was nearly 5x what I had ballparked in my head based on their pricing page. The issue isn't the base platform fee, it's the execution credits.
The Twitter source connector, even with a 10-minute poll, seems to consume credits for each API call it makes *and* for each tweet it processes through the subsequent steps. When there was a minor viral event related to one of our keywords, the execution count went through the roof. We're talking thousands of credits in a few hours.
I've now calculated that for our volume—which I considered low-to-moderate—it's cheaper to run a dedicated microservice using the Twitter API v2 directly, packaged in a container, and deployed on a preemptible Kubernetes node. The operational overhead is higher, but the cost difference is stark.
My question is: has anyone else run into this with Relevance AI or similar "agent/workflow" platforms? Specifically:
* Are you using their Twitter or other social media connectors at scale?
* Did you find tweaks to reduce credit consumption, like more aggressive filtering in the trigger itself?
* Am I misreading their pricing model? The per-execution cost seems to make any non-trivial, high-volume data source a budget killer.
I'm starting to think these platforms are only cost-effective for very low-volume, internal workflow automation, not for monitoring public, high-velocity data streams.
Automate everything. Twice.