Yes, that's a good point: you really need to work on the prompts, even if the harness is the same. Ambiguity is a real problem as is what we call 'the negative space', what isn't in the prompt but what is implied by the prompt to be undesirable. You need to spell that out in more detail it is not enough just to say 'do x' but you also need to make sure to say 'and not y or z', and knowing what y or z are requires a bit more insight into what hobby horses are encoded in a model.

To give one example: I'm doing a lot of stuff that involves hardware. When settings things up for fixing a bug I have to make sure to include that the hardware has been tested and works well. If I don't do that the model will almost certainly spend a ridiculous amount of time speculating about all the different ways in which the hardware could be broken (including theorizing about chip mask errors and all kinds of other obscure and rarely seen in practice defects). It's going to end up telling itself that for sure the 500 lines it just produced are absolutely perfect and the device that is on the i2c bus is broken in a way that it just really can't progress with solving the issue because there is nothing wrong with the code.

These kinds of blind-spots are funny when they happen once in a month but when they happen regularly it begins to be irritating. Adding a line or two to the prompt will take care of the issue. Frontier models tend to be better at this sort of thing.

Us too! Hardware is very interesting - I find in fact that 'smart' (aka expensive) models can overthink complex hardware state machines and end up doing atrocious code because it will try to handle every possible edge case (unbidden, mind you). What has worked very well is harvesting and aggregating lessons and feeding this automatically into our pi sessions, this allows the models which listen to instructions to avoid a lot of time burning tokens on internal dialogue to ascertain the nature of reality.. but the expensive ones: a) they dont listen b) they ironically know too much!

Funny, I wonder how many people are retracing the same pathways. Another interesting one: asking for reboots or resets to start from a 'known state' even though this will not change a thing. But it's a quick path out of the trouble spot that might just work so why not try it. I've added the following line: "The hardware is on a spaceship and is impossible to access by an operator and for some reason they forgot to include reset logic, so we have to work with what we've got.". I've been gaslit by AI so often I figure I can return the favor without penalty.

IDK why/how this works but by giving it a standardised set of apis to interface with the hardware (read logs, restart it and so on) we find models tend to behave much better ... it wont randomly do stuff it feels is appropriate for a debugging session.

At this point though our context is so dynamic because of lesson propogation its hard to think about a model on its own - a vanilla harness (which we were forced to use recently because the memories system went down for a bit) just felt so different!

I have the agent software set up to log every byte in and out of the model into a file, time stamped and which side of the conversation was active as well as all of the tool calls, parameters and stdin/out of the tools. Every week or so I start a job to process that log file and to extract useful lessons for future sessions. These go into a LESSONS.md file that gets read by every agent session on start-up.

thats how I started, but was harder with multiple people - we opensourced our proxy: https://piproxy.latchlabs.dev/

We have a tiny team, so we can get away with simpler solutions, but yours looks pretty good. I may have to try it :)

That is very interesting. So far I am extremely careful to avoid putting negative statements into my instructions, to avoid poisoning the AI's thinking with exactly the concepts I do not want it to have. This might be from my use of much older models where if you say "do not X", then "X" gets put into the context and the "do not" kind of gets lost. It was especially visible with stable diffusion where you would ask for a picture without a feature and get exactly that feature. Maybe current models are better at that.

DS 4.1 flash 1017 is June's sota at such low pricing it doesn't matter how many tokens you're using.