How we keep track of who's talking about our brand
People talk about you wherever they like: a news article, a Reddit thread, a tweet, a forum post weighing you against a rival. On most days none of it needs an answer. Then one day a single angry post with a big audience needs a reply within the hour, and your team only hears about it after everyone else already has.
We hand this watch to an agent on Machine. Once a day it scans the web for fresh mentions, judges the sentiment of each, and posts a roundup in our marketing Slack channel. It doesn't reply or post anywhere on its own. When something matters, it writes a suggested response and leaves the sending to a person.
The problem
Mentions don't stay on one platform. They turn up in press coverage, social posts and forum threads whenever they please, and most are harmless: a passing name-drop, a neutral comparison, a share. The few that count look just like the rest until somebody reads them. And by the time anyone gets round to reading, a nasty post with a wide reach has usually had its most damaging hour.
The common fixes leave gaps. A Google Alert triggers on keywords and knows nothing about tone or audience, so a five-word tweet and a much-shared bad review arrive looking identical. Checking by hand finds yesterday's mentions today, assuming someone remembers. And neither keeps track of what you've already seen, so tomorrow the same viral post gets flagged all over again.
What we built
A daily timer nudges the same long-lived agent session awake again, instead of starting from nothing. On each run it searches news, social platforms and forums for new brand mentions, compares them with a ledger of mentions it has already reported, and judges the sentiment of anything fresh. Then it posts a roundup to Slack: a one-line tally for the routine mentions, and a full write-up with a draft reply for anything important or negative. It never posts, replies or comments anywhere else. The Slack roundup is all it produces.
How it works
Wake the same session every day
A cron trigger (a timer) runs the project once a day, and it reopens the same session each time rather than spinning up a blank one. Since the session carries over, the agent remembers what it reported yesterday and shows you only what's new today.
Tell the agent what to listen for
The brand name, its aliases, product names and the usual misspellings are kept in skills and memory, the playbook the agent keeps and improves. The running ledger of every mention it has already seen lives there too. When a new product line launches or people start using a new nickname, you update the watch list the same way.
Let it search and read the open web
These are public pages, so the agent uses the sandbox's built-in web search and fetch: no connector, no login, no credential. A narrow Slack connector handles the roundup. With that, the agent can:
- Search news, social sites and forums: for the terms on the watch list, for this run only.
- Open each likely mention in full: enough to judge sentiment and reach, rather than relying on a search snippet.
- Post in Slack: a single roundup each run, in the marketing channel.
Keep replies in human hands
The agent searches and reads public pages and writes to a single Slack channel, nothing more. It never replies, comments or publishes anywhere on the open web. For anything flagged as important, it writes a suggested response and labels it clearly as a draft waiting for a person's approval. Any credential it needs is stored encrypted in the Secrets Manager and only handed to the agents you grant them to.
Read one roundup a day
Every run finishes with a single Slack message: how many new mentions turned up, a split by sentiment, and, for anything negative or important, the source, the quote, the reason it was flagged and a draft reply. Routine mentions get a one-line tally instead of the full treatment. Your team reads one message and decides whether to say anything back.
The pattern
A daily cron wakes one reusable session. It searches the web for brand mentions, skips anything already in its memory ledger, judges sentiment and posts a Slack roundup. Important mentions come with a suggested reply that is drafted and never sent.
Guardrails
This agent listens to public talk about your brand, so it can only search and read, and it has just one small way out:
- Isolation. Each run happens in its own sandbox, and the only thing that leaves it is the Slack roundup it is explicitly allowed to send.
- Scoped secrets. Any credential the agent uses is stored encrypted in the Secrets Manager, delivered into the sandbox when the run starts, and only handed to the agents you grant them to.
- Draft, never post. One Slack roundup per run is the agent's only action outside the sandbox. It never replies, comments or posts on the site where the mention appeared. A suggested response is a draft that a person reviews and sends themselves.
- Everything is code. The brand terms, what to watch and the rules for sentiment and importance are files in your repo. They're versioned, and you change them through a reviewed change request, not a dashboard toggle.
The outcome
Your team stops learning about bad press from someone else, and nobody reads yesterday's viral post twice because it got flagged again. The channel gets one roundup a day, quiet when things are quiet and pointed when a reply is needed. Every response that goes out is one a person decided to send.
More use cases
How we answer GDPR data-subject requests before the deadline
Our Machine agent checks each GDPR access or deletion request, finds the person's data across our product database, builds the report well within the SLA and hands it to legal before anything is sent.
How we keep our docs matching what the code does
Once a day a Machine agent reads every commit merged since its last pass, finds the doc pages those changes made wrong, rewrites them and sends you one pull request to review.
How we test every pull request without lifting a finger
Our Machine QA agent works with GitHub and our test environment. For each pull request it checks out the branch, runs the tests, tries the change for real, and posts what it found.
Put your first agentto work today.
- SSO
- RBAC
- Governed agents
- Full audit trail