The feedback you act on is the feedback someone forwarded you. A Reddit thread from three weeks ago, a reply under your launch post. The rest sits there, and the person who wrote it has already given up on you.
Reading it by hand is a chore you’ll do once. So hand it to an agent that already runs on a schedule. We use Hermes Agent, Nous Research’s open-source agent, which can run a job as a cron: ours has a GitHub account and calls TaskFuel to search Reddit and X, then opens a pull request when a complaint describes something it can fix in code.
What the agent reads
The X search guide and Reddit search guide describe the search operations. For the agent connection, start with the Hermes setup guide.
The search is the part your agent can’t do on its own today, and it’s the part TaskFuel sells. Two calls cover the public web where people actually complain.
Search Reddit and X for Linear and tell me what's worth acting on.
- Searched Reddit for "Linear app": 5 posts, $0.02
- Searched X for "linear.app" or from:linear, no retweets: 20 posts, $0.005
- Read all 25 posts
Most of it isn’t a complaint. There’s a Linear vs Jira comparison on Reddit, a product announcement from @linear, and someone explaining why they like the app:
@AdrianDemian_
For me, Linear is the reference. Their design principles explain why navigation should recede while the work gets the attention.
One reply is about a beta setup step people get stuck on. That’s the one worth acting on:
@SophiaChen2003
@linear For anyone joining the beta: the page says GitHub with Diffs already set up on desktop, plus iOS or Android 1.100.0 or later. That desktop setup is an easy prereq…
Paid from your TaskFuel balance · $0.025
Both prices are what I was actually charged.
You do not need a query clever enough to find the complaint, because the agent reads every post and tells you which is which. That judgment is the thing a keyword search was never going to do for you.
Every post quoted above is real output, and the chat shows only part of it. The Reddit search returns the title, author, subreddit, score, comment count, full selftext and a permalink, and the X search returns author, full text, likes, replies, retweets and timestamp, newest first.
Those people wrote for free, because they wanted the product to be better. Most of these threads never reach a roadmap.
The loop
The pieces are ordinary. A Hermes cron job runs a script once a day. The
script calls TaskFuel for the search, hands the posts to the agent, and the
agent decides what to do with them. GitHub access is the agent’s own gh
session, signed in as a bot account.
07:00 search Reddit + X for the product name
07:01 read the posts. keep the ones that describe a fault
07:02 for each fault, ask: is this in the code, the copy, or the product?
07:03 in the code → read the repo, fix it, open a pull request
in the copy → edit the docs or the error string, open a pull request
in the product → post it to an issue, labelled, with the quoteThe last line is the important one. Most feedback is a judgement call about what the product should be, and an agent has no business deciding that. A tag change on a button, a missing heading, an error message that reads like a stack trace: those it can fix. “I want offline mode” it files and forgets.
A pull request is also the right output shape, because it is the one GitHub artifact that assumes a human says yes. The agent does the reading and the typing. You do the approving.
You run once a day for a product called PRODUCT and its repository at github.com/ORG/REPO.
Search Reddit and X for recent mentions of PRODUCT using the taskfuel CLI, and read the full text of every result. Work out which mentions are feedback: a complaint, a question you can answer by fixing something, a fault.
For each piece of feedback, decide whether it describes a fault in the code, a fault in the copy, or a product judgement.
For a fault in the code or the copy, find the cause in the repository, make the smallest change that fixes it, and open a draft pull request that quotes the post it came from and links to it. For a product judgement, open an issue with the quote and label it needs-decision. Do not open a pull request for a product judgement. Do not act on the same post twice.
What this costs
A day’s search is two calls. About 75 cents a month at these prices, which is less than one coffee for a feedback loop that runs while you sleep.
Searching every day is also how you get the fresh ones. A post from yesterday has an author who still remembers your product. A post from last month has an author who moved on.
Search for the product name and nothing else. You are not trying to write a query that finds complaints, because most posts about your product are not complaints and you would miss the ones that do not use the word you guessed. Read everything and sort it yourself. Cap it at a couple of pull requests a day and let the rest queue as issues.
When TaskFuel is worth it
Your agent can’t do the reading part on its own. We asked Claude Code, without TaskFuel, for recent Reddit and X posts about Linear. Its fetch tool is blocked from Reddit, X answered with a 402, and web search came back with a vendor blog summarising Reddit rather than a single thread.
Monitoring tools sell this by the month (prices checked October 2, 2026). Syften starts at $29.95 for Reddit, X and Hacker News, and Brand24 at $249. If Reddit alone is enough, F5Bot emails you keyword matches for free and is the better deal. The two daily searches here come to about 75 cents a month, and an alert still leaves the reading to you.
By hand, it’s the same two searches and all the reading, every day, done by you. The cron job does it at 07:00, before you’re up, and hands you only the pull requests.
Connect your agent and the first $1 is on us.