5h ago

Chris Anderson argues decentralized agent markets are stalling because models cannot estimate their own token usage or costs

Agents cannot price bids effectively without baseline resource forecasts

——0——
Original post
OPChris Anderson#1839Chris Anderson@CHR1SA

Agents suck at estimating how many tokens they'll use and thus how much they'll cost to run, according to this fascinating paper from @krishnanrohit. This explains why agent marketplaces have been so slow to take off. https://andreyfradkin.com/assets/marketbench.pdf?utm_source=substack&utm_medium=email From latest @windfalltrust brief https://windfalltrust.substack.com/p/brief-11-mismeasuring-productivity?utm_source=substack&utm_medium=email&utm_campaign=email-restack-comment&r=7khick&triedRedirect=true

11:52 AM · May 26, 2026 View on X
REPLYrohit#1220rohit@KRISHNANROHIT

@chr1sa @roybahat @windfalltrust Yup!

More that without a better estimation, at least credible on an ongoing basis, it's hard for models to figure out how to bid for a task. As long as they do sensibly, it's fine even if they're not accurate per se.

9:58 PM · May 26, 2026 · 12 Views
REPLYrohit#1220rohit@KRISHNANROHIT

I think this is true, but the key point is, if you have a dozen models and a dozen scaffolds, how do you choose where to route the question. Well the models should be better calibrated to say "this is what I'll bid". Will be worse for completely unseen problems but need to be good enough to engage in a market (as we do as humans).

10:31 PM · May 26, 2026 · 29 Views