Case studies · AI workflows
An AI research team that works the night shift.
A trading research platform tests strategies that are promoted online against real market data and real trading costs, and publishes every verdict, including the ones that don't hold up.
The research, testing and publishing run as a pipeline of scheduled AI agents, each with one job, hard limits and a human sign-off before anything goes public.
By the numbers
- 41
- strategies in the pipeline so far: 25 published, 12 rejected
- 6
- AI agents on a nightly or daily schedule, with 5 more roles planned
- $0.50–$2
- typical cost of one agent's shift, with a usage cap on every run
- 100%
- of agent changes audited before anything is saved
How the pipeline works
A simplified view. Each step runs on its own schedule and hands its work to the next.
Overnight
1Gather
Research agents
Kai and Ada each bring in one idea a night, written up in a fixed format. Two more researchers are planned.
Overnight
2Check
Vera, independent checker
Re-checks every idea her own way, never with the researcher's work.
Does it hold up?
Most ideas stop here.
Rejected
Logged with the reason, so it can be revisited.
Survivors only
3Build and test
Built and tested
Automated end to end, one step at a time.
Review gates
Each step is signed off before the next. Nobody approves their own work.
Morning
4Approve and publish
Write-up drafted
The verdict and the evidence behind it.
Remy, Sage and Gia
Marketing agents for content, search and analytics.
Owner approves
Nothing goes live, gets posted or costs money without a person signing it off.
Live on the website
Published.
The controls around the agents
This is the part that matters for a business. Agents that can act on their own need limits that hold even when a prompt goes wrong.
- Each agent has a job description and a fence
- Every agent has a written role and its own list of what it may and may not touch. Requests outside that list are refused and logged.
- Spending and time limits on every shift
- Each shift runs with a usage cap, usually $2, and a wall-clock limit. Recent shifts took one to eight minutes and cost between $0.52 and $2.03.
- An audit before anything is saved
- After each shift the runner checks every file the agent changed against its policy. If anything is out of bounds, nothing is committed.
- Rules enforced in code, not in prompts
- The most sensitive steps are locked in code and won't run without explicit approval, whatever an agent asks for.
- Nobody checks their own work
- Every result is re-checked independently by a different agent, and build steps pass separate review gates before they count.
- A person signs off what matters
- Anything going live or being posted, any paid spend and any new market data needs the owner's approval. Every shift writes a plain-English report.
What it's built on
Built and run by our lead architect, end to end.
- Agents
- Claude Code running headless on a schedule, one profile and permission set per agent
- Automation
- Python automation driving desktop trading software that has no API
- Website
- Next.js on Vercel, preview builds for every agent change
- Data and reporting
- GA4, Search Console and the email list pulled into a daily report
The same pattern, on your data
Swap trading strategies for supplier quotes, support tickets, compliance checks or sales leads and the shape is the same: agents gather and draft, an independent step checks, rules are enforced in code, and a person approves what leaves the building.
We can build that around your systems and data, starting with one workflow at a fixed price, so you can judge it on something small before committing to more.
Got a process that runs on people copying things around?
Tell us what it is. We'll say honestly whether an AI workflow would help, and what it would take to build one safely.