I built agents to do humanities research, social media research, build websites, and write daily briefings. They did useful work. I wanted to keep that work, run it again, and improve it.
That is where I kept getting stuck.
A report would improve, but the instructions for tomorrow’s report would stay the same. An agent would save a debugging guide, but I would have to explain the procedure again. A failure near the end of a long run would mean repeating steps that had already finished.
I asked Codex to search six months of my local conversation history. It searched 147,968 message records and read the surrounding exchanges for the strongest examples.
Here are five.
What did you actually change?
I had a daily automation that reviewed my work and wrote a briefing. One report contained pasted error-page text where a readable summary should have been.
The agent cleaned up the report, changed the code that displayed it, and deployed the result.
I asked:
What did you actually change? Did you ipdate the prompt for the automation? Remember that’s what this “is” an an automatipn. Show me the prompt.
The agent replied:
I changed the renderer, not the automation’s task prompt.
It then inspected and updated the automation prompt. That change had not been part of the repair it initially reported as complete.
I still see the same prompt
After several failures in a dataset-maintenance automation, the agent simplified the process. It reported that the automation now used a single command and that the old instructions were gone.
I opened it and asked:
Great. Are you sure it works ? Did you update the codex automation? I still see the same prompt in that daily automation.
The agent checked the file and said it contained the new prompt. It suggested I might be seeing instructions from a previous run.
I pasted what I could see.
After another edit and inspection, it found the mistake:
I see the problem: my patch added the short prompt but did not remove the old prompt = ... line, so the file now contains two prompt keys.
It removed the old entry and checked that only one remained.
I keep having to remind you how to do that
March 18 and 26, 2026
We had worked out how to retrieve the logs for a failed research session using a session cookie saved in the project.
I asked the agent to record the procedure:
Update your guide that tells you how to debug a session, that tells you where this session cookie is, so that I don't have to keep telling you the same thing.
The agent reported that it had updated and saved the guide.
Eight days later:
What I want to be able to do is be able to just tell you anytime something goes wrong to check the logs. When you check the logs, you check the entire log bundle. I keep having to remind you how to do that.
The agent found the existing guide and expanded it. I also asked it to save the working script so it would not reconstruct the request each time.
The guide existed. The record does not establish whether each later agent read it.
Why does it take two hours every time...?
I was building a generated wiki. Deployment failed because a search file was too large. The agent split some of the data and tried again. Another file still exceeded the limit.
I asked:
Did it work? Why does it take two hours every time to try this?
The collection and analysis had been saved. But each retry still repeated content generation, the site build, search-file generation, and deployment. The agent estimated another one to two hours.
My response:
Save these so you don't have to redo them every time. And yes, split those, do whatever you need to do to make this work, and to not make it take two hours every time.
The file-size limit was an ordinary engineering problem. Repeating finished work made each repair expensive to test.
Why did you stop working?
I was building a system that converted a screen recording into workflow instructions.
Parts of the system passed their tests. My actual recording still failed. A filename was present in the extracted screen text, but the agent omitted it from the record used by later steps. A check caught the omission and stopped the build.
I asked:
Why did you stop working ? Do you only analyze 36 of the frames ? Is that enough ? Why is the goal marked blocked ?
The agent replied:
You are right. I should not have stopped after I found a fixable production failure.
It also found:
The goal record has no stored reason for the blocked state. I cannot justify that state from the record.
A later acceptance test required access the agent did not have. But it acknowledged that it could still repair and test the earlier stages.
Why I’m building Method
Improving these agents often meant reconstructing the previous attempt: finding the instructions, checking the results, and working out whether a correction had been saved.
I’m building Method to make those things available to inspect. A Method is a workflow file with instructions and checks for each step. Each run keeps a separate record of its results.
You can review the changed instructions, run the workflow again, and inspect whether the repair worked.
