October 2026 AI Papers, Applied to Your Agent Setup

Four of October's papers can be applied with a text editor.

We have already covered this month twice. Our notable AI research papers of October 2026 says what seven papers found, and what would change our picks says what evidence we need before any of it moves a tool recommendation. This page skips both questions and goes straight to setup: the exact lines to put in an instructions file, and the habits to change in a session, if you want to borrow the results before anyone replicates them.

I read the pages ranking for this month's papers before writing. Two are daily digests of three to four hundred words each, sorted into themed buckets like generative models, computer vision, and multimodal reasoning. The third is a yearly index of about 800 words that mostly lists titles. They cover what came out. Not one of them contains an instruction a reader could follow. So that is this page.

One warning first. Every paper below is a preprint, and every figure is the authors' own, taken from the arXiv abstract or DAIR's weekly summary as reported in our roundup. Nothing here has been benchmarked by us. These are cheap experiments, and you should treat them that way.

The short answer

Harness², which lets an agent rewrite its own harness without a ground-truth verifier, is left off on purpose. It pulls the opposite way from the third row, and I would rather you have a grader than a self-editing loop.

Restart the session when the diffs shrink

The Berkeley, Stanford and Bespoke Labs team found that agents stall by anchoring on their own past attempts, making fewer and smaller edits over time. Wiping the history and restarting from the best result so far gave up to 39% over a single agent and up to 15% over parallel agents on the same budget.

The signal to watch for is the diff. When a Claude Code or Cursor session has spent forty minutes on one function and each change is a line or two smaller than the last, it is anchored. Do this:

1. Save the best version, as a commit or a copied file.
2. Write two or three lines saying what that version does and what is still wrong with it.
3. Close the session. Do not compact it, because a compacted summary carries the same anchors in fewer words.
4. Open a fresh session, point it at the file and your three lines, and nothing else.

Step three is the one people resist.

Commit the dead ends instead of resetting them

GitSwarm had hundreds of asynchronous agents share one Git repository, with every contribution landing as a commit. DAIR reports that 70.1% of failed research commits were cited by later accepted work. The failures helped because the next agent could read them.

Most of us do the reverse. A Claude Code attempt goes wrong, we git reset --hard, and the reason it failed lives only in a closed terminal. If you run Claude Code and Codex on one repo, both can read the same rule, as long as CLAUDE.md is the single line @AGENTS.md, the setup our Claude Code vs Codex comparison walks through. The rule I would add:

## Failed attempts
- When an approach fails, commit it on a branch named `attempt/<short-topic>`
  before reverting. The commit message must say what was tried and why it failed.
- Before starting a fix, run `git branch --list 'attempt/*'` and read the
  messages for anything related.

Branches keep the dead ends out of your main history while leaving them findable. And keep the rule short. Codex caps the combined AGENTS.md it reads at 32 KiB, and this site's own file is already 8,424 bytes.

Give the agent a grader it cannot argue with

The Harvard, QuEra and Anthropic paper, DAIR's Paper of the Week, had an LLM propose reversible circuit designs while a separate evaluator verified and scored each one. The discovered circuits used fewer Toffoli gates and fewer qubits than the state of the art in seven of eight tasks. The model never graded itself, and the same was true in GitSwarm and the forgetting paper.

Your coding agent already has a version of that evaluator if your repo has tests. The gap is usually that nobody told the agent which command counts. Claude Code runs tests when it knows how, and Codex in its default workspace-write mode will run npm test without stopping to ask. So name it:

## Definition of done
- A change is not done until `npm test` and `npm run build` both exit 0.
- If either fails, fix the cause. Do not edit or skip the test to make it pass.

The second line matters more than the first. A grader the agent can rewrite has stopped being a grader.

This repo learned that one the hard way. Its AGENTS.md records that the "Title too long" issue was filed twice, because the first fix was checked by hand and shipped no automation, so nothing kept it closed. Now an audit fails the build when a title runs over budget, and the instructions say plainly not to downgrade it to a warning. That is the evaluator pattern, written by people who had never read the circuit paper.

Split what you found from what you concluded

LongHarness names two failure modes anyone using a research assistant has met. "Unbounded exploration" keeps retrieving without closing the gap that matters. "Evidence-state drift" is when a fact the agent already found, or a correction you gave it, stops shaping its answer. Its fix keeps evidence and conclusions as separate, explicit state, and on OfficeQA Pro it lifted GLM-5.3-Flash from 15.8% to 60.2% exact match.

You cannot install LongHarness into NotebookLM or Claude Projects. You can copy its bookkeeping by hand. Keep two short documents in the project, one called evidence (sourced facts and your corrections, each with the file and page it came from) and one called open questions (what is still unresolved). Paste both at the top of each new chat, and ask for the answer to cite the evidence list.

It matters most in Claude Projects. Our NotebookLM vs Perplexity Spaces vs Claude Projects test found that past the 200,000-token context, Claude switches a project to retrieval mode and reads only what its search pulled back, with users reporting the switch around a dozen files. Two short notes stay small enough to be read in full on every turn, even after everything else gets retrieved.

What I would try first

The restart habit. It needs no file edits, it works the same in Claude Code, Codex and Cursor, and it attacks the failure I see most, a session polishing the same twenty lines for an hour. The definition-of-done block comes second, because it costs one paste and you will know within a day whether your agent respects it. The dead-ends rule I would give a month, since its payoff only shows up when a later session finds an old branch and skips a mistake.

If you want help reading the papers themselves, our AI research tools roundup puts Elicit against Consensus.

Frequently Asked Questions

Which October 2026 AI research papers can I apply without new software?

Scaling Discovery Requires Forgetting (restart stale sessions from the best result), GitSwarm (commit failed attempts with a reason), the Harvard, QuEra and Anthropic quantum-circuit paper (give the agent an outside grader) and LongHarness (keep evidence and conclusions separate). All four translate into habits or instruction-file lines.

What should I add to AGENTS.md based on October 2026 research?

Two blocks: a rule to commit failed attempts on a named branch with a message explaining the failure, and a definition of done that names your test and build commands and forbids editing tests to make them pass. Keep both short, since Codex caps the combined AGENTS.md it reads at 32 KiB.

Does compacting a Claude Code session count as forgetting?

Not in the paper's sense. The forgetting result came from discarding the session and candidate history entirely and restarting from the best result. A compacted summary still carries the earlier attempts, so start a fresh session instead.

Are these results proven?

No. All four are preprints and the figures are the authors' own. The changes above are cheap to try, which is the reason to try them now rather than wait for replication.