Coding
0813 tends to get to work quickly.
Once the request is clear, it usually starts making changes instead of giving me a long list of places I should check myself.
That matters more in older codebases than it does in small demos.
A real bug rarely stays in one file. Change a type and you may have to fix its callers. Change an interface and tests, error handling, and compatibility code may all need attention.
0813 is fairly good at following those connections instead of fixing only the line that threw the error.
It still makes things up, though.
I had one case where it replaced an API from a third-party library. The name looked right. The arguments looked right. The usage looked completely believable.
The API didn't exist.
That's the kind of mistake I dislike more than an obvious syntax error because it survives a quick read.
My rule hasn't changed: read the diff and run the tests yourself.
Long Tasks
This is where 0813 has been most useful for me.
Long coding tasks get crowded fast. You have source files, logs, requirements, test output, earlier edits, and new constraints added halfway through.
With older models, I often split that kind of work into smaller chunks myself.
The problem wasn't that they couldn't handle the code. It was that after enough turns, they'd sometimes lose the reason behind an earlier decision.
A path we had already ruled out would come back. A compatibility workaround added earlier would get removed later because the model no longer remembered why it existed.
0813 does less of that.
If something was established earlier in the task, it usually carries that forward. When a new error appears, it's also less likely to immediately throw away the existing approach and start over.
You won't see much of this in a short benchmark prompt.
You notice it after several rounds of edits and tests.
Agent Use
With a coding agent, I care less about how polished the answer sounds and more about whether it keeps moving.
Reading files, editing code, running tests, and checking errors are basic parts of the workflow now.
What matters is what happens after the test fails.
0813 will often look at the new error and keep working instead of stopping and waiting for another "continue" from me.
That saves friction.
But agent performance isn't just a model problem.
Bad tool definitions, confusing parameters, weak error messages, or overly broad permissions can make a good model look bad very quickly.
When an agent starts behaving inconsistently, I now check the tool setup before blaming the model.
Sometimes the workflow is the problem.
Comments (0)