- From: Евген <arsevgen22@gmail.com>
- Date: Sat, 12 Sep 2026 22:29:44 +0300
- To: public-cogai@w3.org
- Cc: Dave Raggett <dsr@w3.org>
- Message-ID: <CAFF7QG=atb6_4qhrjN67HfEakkn+Ppyb+hQEtn118NzMjNcWmQ@mail.gmail.com>
Dear all, Dave asked me to copy my introduction to the list, so here it is. Short version: Evgenii Arsentev, PhD. I lead AI transformation at a telehealth company, and alongside that I run independent, reproducible measurements of how coding agents behave - my own runs, my own data, published openly. Two of those may be relevant to the group. The first is a controlled experiment on how much context an agent should carry across tasks: 36 runs, and the cost curve came out U-shaped, with a plateau in the middle rather than a single sweet spot (DOI 10.5281/zenodo.22699668). The second looks in the other direction - what AI crawlers actually do when they reach a site. Over 15 days I logged 44,005 requests from 20 crawlers. robots.txt was fetched 577 times. llms.txt: zero. Not once. The rest is tooling and standards work. contextburn measures where the tokens go inside an agent run (github.com/arsentev-ai/contextburn, on PyPI, with an MCP server). I have two Internet-Drafts at the IETF, and eight open peer reviews published on Qeios. Where I hope to be useful here: evidence from real deployments, and review of drafts and demos as they appear. Learning continually from few examples is the part that interests me most. It is also where today's agents burn the most money. Best regards, Evgenii Arsentev, PhD https://arsentev.ai ORCID 0000-0002-9120-7298
Received on Saturday, 12 September 2026 19:31:27 UTC