Automated tools drift

Automated tools drift: posts get stale, plans miss their goal, small problems pile up unnoticed.

Without someone reviewing the whole system regularly, quality slowly falls.

We wanted a reviewer that looks at everything every night, says what matters most, and improves the system safely.

A nightly review that changes how the agents work

The growth brain reviews all of our marketing and outreach tools every night. It was built for our own companies, BinaryBits and TheDataHQ, and it is used every day.

The product screen, shown with test data. Real screens use the same layout.
🌙
A nightly review
Looks at what the tools produced and how it performed, and what is working for others in the market right now.
🩺
A diagnosis
Written at the top of a feed, so the day starts with what matters.
📚
Lessons
Short, evidenced lessons the other agents follow from then on.
🧪
Trials
Small experiments to test a change before adopting it.
🛠️
Fixes it can make itself
And code changes it prepares for a person to approve.
🐞
An issue tracker with AI investigation
Problems noticed while using the tools are logged. An AI investigates each one and finds the real cause: the tool, the settings, the data, the platform, expected behaviour, or "needs more information".

Review, learn, improve safely

AI agentAutomated stepPersonData or serviceWhat the tools didand the resultsMarket researchwhat works right nowGrowth brainnightly reviewDiagnosis, lessonstrials and fixesAll other agentsfollow the lessonsA person approvescode changesLogged issuewith saved contextIssue reviewerfinds the real causeCause & next stepPerson or code fixevidencemarket signalswriteslessonsto approveissueone causenext step
The brain changes guidance and runs trials. Anything that needs code goes to a person.
1
Gather the results
Everything the tools produced, how it performed, and what is working for others in the market right now.
2
Diagnose
The brain writes a diagnosis at the top of a feed, so the day starts with what matters.
3
Write lessons and start trials
Short, evidenced lessons the other agents follow, and small trials to test a change before adopting it.
4
Fix what's safe
Guidance and settings it is allowed to change, it changes.
5
Prepare code changes for a person
Anything that needs code is prepared and waits for a person to approve it.

Reasoning from the goal, not a checklist

Lessons need evidence, and anything that needs code goes to a person. The agents run on Anthropic's Claude, through Claude Code skills.

WhoWhat it does
Growth brain (AI)Reviews the tools' output and results each night, diagnoses, writes lessons, starts trials and makes safe fixes.
Market researcher (AI)Finds what is working for others in the market right now.
Issue reviewer (AI)Compares the output with the context saved when an issue was logged, then names one cause and the next step.
A personApproves any code change the brain prepares.
Learning without copying

Lessons describe principles, not examples to copy, so output stays varied and keeps improving.

Evidence first

Every lesson needs evidence, and issues are traced to one cause, which means fewer false fixes.

Safe self-improvement

The brain changes guidance and runs trials. Code changes need a person.

Clear priorities

A diagnosis at the top of the feed says what to fix first.

Technologies used

Our own Python code and AI agents built on Anthropic's Claude.

Python Claude agents

Challenges and how we solved them

❌ Challenge
Learning without copying.
✓ What we did
Lessons describe principles, not examples to copy.
Result: Varied output that keeps improving.
❌ Challenge
Noise vs real problems.
✓ What we did
Every lesson needs evidence, and issues are traced to one cause.
Result: Fewer false fixes.
❌ Challenge
Safe self-improvement.
✓ What we did
The brain changes guidance and runs trials; code changes need a person.
Result: Improvement without risk.

A review every single night

Nightly
Runs every night
Recorded
Lessons and fixes
1 cause
Named for each investigated issue

The growth brain runs every night, and its lessons and fixes are recorded. Business results will be added here once they are measured.

Last updated: November 2026

What this means for you

Any team running several automations benefits from a reviewer that never skips a night.

  • Self-monitoring for your AI workflows. A nightly review of what your agents produced and how it performed.
  • Automated quality reviews. Evidence-based lessons your agents follow, tested with small trials.
  • An AI that tells you what to fix first. A diagnosis at the top of the day, and issues traced to their real cause.

For client systems we build on PostgreSQL, MySQL, MongoDB or Redis, and cloud data warehouses where needed, chosen for the project's scale and team. See our AI agent development services.

Frequently asked questions

Can it change code by itself?

No. It prepares changes; a person approves them.

What does it learn from?

Results of the tools' own work, and what's working in the market right now.

How is an issue's cause decided?

An AI compares the output with the context saved when the issue was logged, then names one cause and the next step.