Claude Code’s product lead talks usage limits, transparency, and the “lean harness”

"We have no grand plan," says Anthropic's Cat Wu—but that's by design.

See full article...

 

[Speaking Sirloin]

When on the exponential, of course many ideas have to be dogfed to build where the puck or pack is going. But how can an un-opinionated harness manifest a futureproof for every dog?

Edited for clarity.

 

[IronHam]

When on the exponential, of course many ideas have to be dogfed to build where the puck or pack is going. But how can an un-opinionated harness manifest a futureproof for every dog?

Edited for clarity.

I think pi.dev is an example of an unopinionated harness that conforms to its user. Of course, at token cost.

 

[Walternate]

The bite is that shortly after announcing the usage limit increases, they then cut off those creating their own harness (for opinion or just management of all those free-floating tabs) from the Claude Code subscription, by making SDK access API only.

I can see why, if people are kicking off hundreds of subagents (thanks clawdbot), but you’d imagine there would be a better way to control what is happening in your own system. They have rate limits, why not spawn limits, or something, if that’s the problem?

The other shitter is that, oh look, we’re all lining the fascist oligarch’s pockets once again.

 

[ChronocidalManiac]

Would be interesting to see whether they view 'tokenmaxing' as exploration or inefficiency and how they feel that will play out. Also, whether they are committed to addressing controversy around the Memphis datacenter.
I'm a fan of the efficiency gains from AI coding especially Claude, but it's clear the chagrin and shame will be worse if we are not open minded enough to question ourselves often and early.

 

[olumpera]

Amusing to read one article about token optimization at Claude and another one about "tokenmaxxing" at Amazon at almost the same time.

 

[Crito]

The silent change in cache storage default from 1 hour to 5 minutes made a notable difference in our enterprise expenses. Especially when running parallel tabs it is very unlikely you’ll get back to the blocked tabs in time.

I had a few users A/B test a beta flag to switch it back to 1 hour (now at 2X the cost per input token) and it still resulted in a cheaper experience (20-40% reduction across all users). Pretty major/breaking change to be deploying without notice, but that’s par for the course at this point.

I really love Anthropic and think most of the people who work there are great, they've just expanded ridiculously quickly and enterprise is not used to the pace of change (nor should it be).

Source: I partner on admin-ing Claude at a T10.

 

[arakon]

I remember a time that not having a plan for something related to business was considered a ‘bad thing’ that would get your funding pulled. They sure have gotten good at spouting words at you though. Making up new buzzwords to describe your process without actually defining them is peak un-transparency.

 

[jhodge]

So it's all sort of ad-hoc and demand-driven at the moment because the fit between what they can deliver and what users want is shifting too quickly for long-range planning. OK; makes sense. OTOH, I imagine that sooner or later "it needs to pay for itself" will serve as a disciplining function.

 

[soybot]

it's really hard to figure out where the tokens are going? That seems like a fundamental flaw in a supposed money making enterprise!