I check my positions too often. I always have. The cure isn’t willpower — it’s automation.

Every weekday at 21:00 Europe/London (after the US market closes), an OpenClaw cron called market-watchlist fires. It pulls prices for 7 tickers I care about, formats them as a clean message, and only does the expensive thing — deep research — for tickers that moved more than 4% that day.

That last constraint is what makes this work. Without it I’d be paying for 35 deep-research runs a week to learn that nothing happened.

The structure

The job runs 7 sub-agents in parallel, one per ticker:

• TSLA, NVDA, AVGO, META, GOOG (equities I hold or watch)
• BTC, ETH (crypto)

Each sub-agent does the same simple thing: hit Yahoo Finance, get today’s open, current, change %. They report back to a coordinator agent.

The coordinator does two things:

1. Builds the daily summary message (always sent)
2. For any ticker with |change%| > 4, spawns a research sub-agent that pulls news headlines from the last 24h and summarises why the price moved

A normal day: just the summary. A noisy day: summary + 1-3 research blurbs explaining the moves.

Why 4%

Because below that, “news” is just noise. A 1% move on TSLA on any given day means nothing. A 6% move means there’s a story I should know.

I picked the threshold by eyeballing about a month of data. Below 4%, the “research” was always “general market sentiment” or some analyst saying nothing. Above 4%, there was usually a real reason: earnings, regulatory news, a tweet from someone who shouldn’t tweet.

What this looks like as code

The whole thing is one cron entry. Conceptually:

spawn 7 sub-agents [pull price] in parallel
  └── coordinator collects results
        ├── format daily summary
        └── for each ticker where |change%| > threshold:
              spawn research sub-agent
              └── append research to message
send message to Telegram + Discord #finance

It runs on the local Qwen3.6 model. Each ticker pull is one HTTP request. The research sub-agents do a web search and a summarisation. End-to-end: about 90 seconds on a noisy day, 15 seconds on a quiet one.

The pattern for businesses

Swap “tickers” for anything you watch.

• Customer support: ping me when ticket queue grows by more than 20% in an hour
• Sales: deep-research any deal that’s moved between stages today
• Operations: alert me on any SKU whose stock has dropped below the 7-day average minus 2 standard deviations
• Site reliability: investigate any endpoint with p99 latency above the rolling baseline

The shape is identical: cheap polling at high frequency, expensive analysis only when something interesting happens. The threshold is the magic. Without it you either drown in noise or miss the signal.

What I’d warn you about

Don’t set the threshold from gut feel. Run the polling for two weeks first, log everything, then look at the distribution and pick a threshold that catches the genuinely interesting cases without flooding you. The threshold should change every six months as your baseline shifts.

Build the kill switch first. My market watchlist nearly DM’d me 47 times in a single hour during the March crypto wobble. I now have a hard cap: max 3 research blurbs per run. Anything beyond that gets summarised into one paragraph.

Decide who reads it. A morning brief I read in bed is one thing. A real-time alert that wakes someone up at 3am is a totally different design problem.

---

This is the same architecture that would watch your business’s noisy systems and tell you only what matters. Tell me what your team is monitoring manually and we’ll see if there’s a 4%-threshold version of it.