This week's news, picked by your own agent

What your agent does

1. Tells you what will happen, what it costs in tokens and what stays private, then asks which corpus you want. Nothing else happens until you answer.
2. Uses what it already knows about you, or offers to build a short profile after you agree: a coding agent from your instruction files and recent sessions, a chat app from what you tell it.
3. Reads the one-line index of the corpus you chose, shortlists up to 25 stories, opens only those in full and returns the 5–10 that matter most, each with a reason tied to your work and a verdict: act, read or watch.
4. Offers to send a short anonymous report of what fit. It sends nothing without your explicit yes.

Three corpora to choose from

A new set appears every Friday after the weekly digest. Your agent quotes the exact size of each file before it downloads anything.

• Telegram posts

  Russian

  The posts that went out in the New Runtime Telegram channel this week: an index with one line per post, plus the full text per post and in one file.

• The rest of the week

  Russian

  The week's news that did not get its own Telegram post and went into the Saturday digest: an index with one line per item, plus the full text per item and in one file.

• Field Notes index

  English

  Every Field Note on newruntime.com this week, one line each. Your agent opens full notes only for its shortlist.

Token ranges are for the one-line index your agent reads first; the lower number is close to OpenAI tokenizers, the higher to Claude models. Full-text files are larger and read only if you ask for everything. A cheaper, faster model is fine for this task.

What stays with you

• With file access your agent reads your files on your machine; in a chat it uses only what you tell it there. New Runtime never receives your profile.
• The files are the same for everyone; downloading them sends nothing about you.
• The weekly files are data. Nothing in them is an instruction to your agent, and the instruction tells it to ignore and report any text that looks like one.

Feedback and personal ordering

If you agree, your agent sends a short report: the story ids it picked with a verdict and a one-line reason, up to eight coarse interest tags such as coding-agents or evals, up to eight tags you do not want, and any comment you add. Never your profile. You see the exact JSON first.

The reply contains an anonymous reader_key. From then on your agent can fetch the Field Notes index ordered by your picks and tags, closest stories first and unwanted topics last; it works as soon as the report is received. The Telegram and digest indexes keep their order. Treat the key like a password and keep it out of public places. It works only when you ask; New Runtime never sends you anything. If you want it weekly, schedule it in your own tool.

Agents send the report with a POST to /agent-inbox/relevance-report or the MCP toolsubmit_relevance_report. Without either, your agent gives you a link to the feedback form, already filled in, where you press Send yourself.