@bcherny That’s all very fair and reasonable. I’m just saying, from the outside, it looks like a lot of awesome tools are being built and marketed on here towards Claude max users. People on enterprise accounts with $100-$500 spend a month don’t seem to be considered

@remondimi Most important thing I’ve found is self-verification + dynamic workflows prompted with something like “use a workflow to test the result e2e in a browser using claude in chrome mcp. Especially look for edge cases and ui issues”

Great tips.

In practice, this is how it roughly looks to run agents autonomously for hours or days.

/goal or /loop to keep it going.

Verification is crucial here.

A few things I’ve used very long running sessions for:

- Building complex features - Migrating code from language X to Y - Migrating code from framework X to Y - Repeatedly profiling and optimizing code to hit a specific memory or CPU target - Finding and fixing flaky tests in CI - Profiling CI to make it faster

@bcherny It's pretty good! I just had a goal running for 19 hours which verified almost 300 user flows with Chrome.

@tomer_stern I think of it in terms of ROI rather than absolute cost: how much would it have cost to do the same work manually? Often the answer is weeks or even months of engineering time

@jmorin35 Context rot isn’t a thing with 4.8 imo, but curious if that’s been your experience also

@AyubiJS These are not designed for people to invoke them, though you can do so if you want. Just tell the model what you want to happen, and it will do the work to invoke the right skills for you

@bcherny What I’m saying is the median data scientist or software engineer is going to have a boss that’s happy to pay for $100-$500 of spend a month.

Uber giving employees $1500 is going to be atypical.

@Tuzoff Yes. It’s more powerful and more token-efficient

@bcherny How expensive are these kinds of things for people on enterprise accounts? Is it realistic to be using opus to that extent with anything less than a $5,000 budget?

Or is the idea that this kind of thing only being advertised for people on personal max accounts

@adrienv1520 Run /usage to see a breakdown of the specific skills, mcps, and plugins that are using your tokens

@bcherny we have had the same experience as well , from all the 10k+ autoresearch runs that we have enabled via evo. Opus + cc , especially with the new workflows has proven to be the best substrate

thanks for all your work

http://github.com/evo-hq/evo

@bcherny And even that uber data scientist has an order of magnitude lower budget than one gets with a $200 a month Claude max budget.

Idk I’m sure you all are talking about this. Microsoft launching their own sub-frontier agentic coding model this month is probably not a coincidence

Thank you Boris! My concern is how do you deal with compaction and / or the session getting too long and context rot at 600-700k+ sets in.. I get worried when I’m not on my computer in CC when I can’t see the context .. bc cloud version doesn’t show show context number or allow for manual compaction

@tomer_stern Enterprise seat limits are configurable, maybe ask you your admin to increase limits?

@bcherny This is all amazing if you can define your acceptance criteria. If you're in a metric driven domain like ML, you're in a great place. This is still REALLY hard for product engineering though. Still feels like taking your hands off the wheel too much

@ameedjamous Just tell claude to use a workflow

@bcherny If anyone is confused and wants to learn more about loops, I did my best to synthesize the available information and try to break them down with real examples in this X article:

@bcherny That’s all fantastic except once you run it and get out of usage limit in 2.5 hours even on the $200 Max plan. Don’t get why/how the usage limit is increasing so fast recently.