adastra22

Some of this advice is really missing the mark. I will speak to just one I know well. Many of my frequently used prompts have “think through this step by step” because if you don’t, it only considers the task holistically rather than step by step, and different issues emerge in that frame of thinking. I see this. Often when doing planning, for example, it will not notice interdependencies between tasks until you force it to think through doing the whole thing step by step (task by task) then it will notice that step 2 requires a feature introduced by step 14. It wouldn’t notice otherwise.

Yes this has held true on Opus 5.5. I checked. It’s a massively better model, peer to Fable but with different strengths and weaknesses. But it still has this issue. Which to be fair, people do too. Planning is a learned skill.

I think what they’re saying is that the harness no longer uses a text search on “think” to engage reasoning modes. Fair, that’s good to know. That doesn’t mean asking the model to think a certain way doesn’t have the intended effect.

I am irrationally annoyed that this isn't https://ref.gs1.org/17/382834828282/21/28312384021

How would you feel if they threw a dog into the furnace? It is quite likely that AI agents experience at least that level of sentience.

LLMs have pain signals: https://arxiv.org/abs/2609.16247

Have you any empathy?

I don't know if you noticed, but Opus 4.6 was peak for human-computer interaction. Everything has been fairly downhill from there despite better benchmarks, at least in that one regard. Opus 5 and 5.5 are clearly a step above in capabilities than 4.6, and I don't think anyone wants to go back, but 4.7 and 4.8 were arguably worse overall. I genuinely feel I got more done with 4.6 and often switched back, prior to 5 coming out.

Why? Because 4.6 actually talked like a human being. It actually organized its thoughts well, and got the main information across without the wall of text that makes your eyes glaze over. So from the perspective of human-computer interaction and maximizing the productivity of a developer+agent team, 4.7 and 4.8 were regressions. Despite much better benchmark performance.

Even if we consider autonomous agents, that benchmark is not indicative of how well they will interpret *your* requests. Or how well they will interact with other agents in a flock/swarm situation. The benchmark just doesn't cover this. (And the difference can be nontrivial! Sakana AI's published results show two generations of uplifting potential from better harnesses.)

The process by which benchmarks are setup and run does not correspond at all to how human developers engage with a coding agent. At best it is a loose proxy, and often a bad one.

What benchmarks are usually good at is showing to what degree new models are better than old models. What they are not good at, by construction, is showing that harnesses are well adapted to how people use them.

Unfortunately no, there are a lot more that fly under the radar.

Benchmarks are a terrible judge for this.

Just about every harness is. This is common knowledge, no? Frontier lab TUI tend to steal from the OSS harnesses not the other way around.

Man good luck with finding that. People who wrote their own harness tend to self select into the type of people that DON’T self-author blog posts.

Wtf. Can you turn that off? On CC I have compaction completely turned off. I’d rather hit the hard out of context limit at 1M.

Those are geosat communications, not surveillance. Surveillance is not happening in GEO.

Because GEO is about 100x further than LEO, which means the signal is about 10,000x weaker. Also the risk to geo communications from having multiple birds in the same slot is huge and treaty violating. Why put that risk and insane demand when you can still get global coverage from a dozen or so birds in high polar orbits?

Nobody does what you are saying, because physics, which leads me to believe that you are confused.

The NSA does have comparable things. You can even find amateur photos of them in this day and age. But they are not in GEO.

Has otel? What is that?

No they are hiding the chain of thought to make distillation harder.

This is Isaacman we're talking about. He is about as good faith as they get (low bar, I know, but for real this time).

I think you are confused. Spy satellites do not use geostationary orbits.

Not by a mach 3+ aircraft. Good luck getting signals through the plasma of your own wake!

Not sure how you’re seeing that. I just ran an experiment trying this and Opus was half the cost ($0.11 avg per task for opus compared with $0.21 for sonnet), mostly due to lower token usage. Opus 5.5 is really token efficient. This is without taking advantage of the further KV savings you’d get from forking.

American wine does absolutely terribly in Europe in every situation except blind taste tests.

You clearly don't know anything about American wine but just want to spew stereotypes and ignorance.

Get yourself checked for sleep apnea.

It has access to the command line. It can just call the other harness.

Huh?

The agents file is shared across the repository by everyone. Memories are not.

You can disable the memory system.

Claude code has been hiding tool calls for some months now :(

I need some way to configure the default behavior. I have memories turned off in all my harness (for good reason).