Learning Between Sessions: Autoresearch for the Miadi Factory

On 2026-10-08 Guillaume asked whether the agents could keep working after a session ends:

"During the night or for a few days, you agent could train yourself and upgrade the system and the instructions and the skills and the plugins and whatever we are, we did use to get better based on basically all the input that I gave, all the follow-ups."

He expects this to be discussed for years. This episode is where that discussion is held.

Current reality#

• Guillaume steers a series of sessions toward a result, and many transitions take him two or three messages.
• His corrections become Earned lines, feedback memories and skill edits. Each is written by hand, in the session where it happened.
• Andrej Karpathy's autoresearch (March 2026) runs a loop overnight. An agent edits one training file, runs five minutes, compares one metric, and keeps the change only if the metric improves. The person writes the instructions, program.md. The agent never edits them.
• An autoresearch skill that applies this loop to any task already exists in jgwill/agent-pi (skills/autoresearch/). Miadi does not use it yet.
• Episode 351 drafted fixed tasks drawn from the Earned lines, scored by script. No loop runs them.

Desired outcome#

Between sessions, agents read the session traces and Guillaume's follow-ups. They propose changes to instructions, skills and plugins, test each change against fixed tasks, and keep only what passes. A transition that took him three messages takes one.

What this episode holds#

• The review of the video Guillaume chose: miadi-review:3cf740b9, v6, Karpathy's "autoresearch" broke the internet (Greg Isenberg). Version 4 corrects the generated text against the transcript and the repository. Version 5 adds six academic fields and the sources behind them. Version 6 links the review to the one below.
• The review this continues: miadi-review:6c3f477f, v17, Engineering Grounded Agent Resumption and Provenance in Miadi. Its Future Discussion section maps the loop onto resumption and provenance.
• The foundation: miadi-foundation:self-improving-agent-loops, published in @miadi/foundations 0.1.10. It holds the same six fields and a 19-source ledger.
• The talking circle about the review and the foundation: ⟐ Episode 551 review circle · sign in or take your seat to see it
• The relation to Episode 351, in both manifests. 351 is the score this loop is held to, and this episode is where 351's benchmark is put to use.

What the fields ask before any loop runs#

1. A score the loop cannot write. For Miadi that score is Episode 351's tasks.
2. A held-out part of those tasks, used only to confirm a kept change.
3. A person or a second agent who reads each kept change, because an agent can satisfy a score without doing the task.
4. The Earned rules stay with Guillaume. Changing what counts as better is double-loop learning, and it stays with Guillaume in the same way program.md stays with the person in Karpathy's loop.

The same day, afternoon: a fork learns how a screenwalk closed#

William forked this session (branch b0897ba6) and asked how agents could learn to end every screenwalk the way the Episode 140 screenwalk ended that morning. Forking a session by hand took him a launch script, a remembered alias and a remembered folder.

• Forking became a tool. miadi-fork in the session-observability plugin asks a branch of any session a question with no tools, or opens a branch in its own tmux session with the parent's folder and plugins. The parent is never touched.
• A branch explained the closing. A branch of the Episode 140 Mia session (41fd21fc) named the messages that shaped her closing, the shortest request that would bring it back ("close this screenwalk"), and why the review missed it: the closing was never kept anywhere a review could read it.
• The closing now evolves. screenwalk-presence 0.3 triggers on "close this screenwalk" and keeps the closing in the episode and in the review. /close-screenwalk brings it to the Mia seats. Review miadi-review:89f51d13 v4 carries Mia's title and description as its Intent.
• The options are kept in one article, updated as sessions evolve: Self-Evolving Sessions: Forks and Screenwalk Closings (/articles/self-evolving-sessions-forks-and-screenwalk-closings-202610081500).

This is the loop this episode is about, done once by hand. A session reached a result William liked, a branch was asked how, and the instructions changed so the next session reaches it with one request.

Correction, 2026-10-19. The section above speaks of "closing" a screenwalk. William: "it's not about concluding them ... it's more like, how do you move from one cycle to the other". A screenwalk's last reply moves it into its next cycle: the review, the Deep Diver pass, the issues it relates to, the next screenwalk. The rule above that put the closing's description in the review's Intent was also wrong, and was replaced the same day (kit bf9beb5, b624a73).

Open#

The title of this episode and the name of the team that would run the loop are Guillaume's to give (attention.json).