Pay for a Mistake Once
- Each project mistake becomes a lesson with a trigger, a fix, and a source.
- Every new session reads lessons from other projects before work begins.
- A written plan must explain how each lesson applies to its stage or why it does not.
- Lessons leave out client names so they can travel across projects.
I direct AI agents across many client projects at once. To carry what I learn from one project to the next, I built a system called HALO. Without that transfer, the same mistake can happen again on another project, costing me time and the client money.
An agent starts each session without memory of the last one. If it needs to know something, that information has to be in files it reads at the start. I treat each session as a fresh start, even when the work continues from an earlier session.
The takeaway
Write each mistake as a lesson, then make every future session read it. A project should pay for an error once.
How do I stop an AI agent from repeating a mistake?
I record each mistake as a lesson and make it available to future sessions. The lesson names what went wrong, the fix, and where the mistake happened. That gives the next project something concrete to check before it starts work.
A chat message does not carry a rule into the next session. A file that every session loads does. That difference changes how I manage the work: an instruction has to live somewhere the next session can read, or it is easy to lose.
A repeated mistake costs time and money each time it happens. The record gives me a way to keep that cost from recurring across projects.
What does a useful lesson contain?
A useful lesson has three parts: the trigger, the fix, and the source. The trigger says what went wrong. The fix says what to do differently. The source says where it happened.
Those parts help an agent apply the lesson without needing the original conversation. "Open every outside link before a page ships" is one real lesson. It gives the agent a specific action to take before a page is finished.
Another lesson says a check is proven only by watching it fail on real input. That makes the evidence for the check clear: a check that has not failed on real input has not yet met the lesson's test.
When a lesson comes from a client project, I leave out the client's name. The cause and fix can still help another project without carrying private client details with them.
How do I make sure a lesson affects the next project?
New lessons reach the next session of every other project automatically. That way, the session can read the lesson before it begins its work, even if the mistake came from a different project.
The written plan for each stage must address every lesson before work begins. For each one, the plan includes a sentence explaining how it applies or why it doesn't. A check rejects any plan with a blank.
This step makes the lessons part of planning, rather than background material an agent may skip. The plan has to account for each lesson before the work begins.
Common questions
Does this mean an AI agent remembers every past session?
No. An agent starts each session without memory of the last one. Information it needs must be in files it reads at the start, so a team should check that its setup loads those files before work begins.
Do lessons from one project expose the client to another?
The lesson leaves out the client's name, so it can travel across projects without naming the client. The lesson records the mistake and its fix, not the identity of the client involved.
How can I evaluate an agent before it handles live work?
Evaluate the agent before it touches production. Check that the work it is meant to do is covered by clear instructions and that its checks are tested with real input.
If you want me to build an agent system that carries lessons across projects, contact me.