The changes from 14–18 October address what reviewers see and where they send corrections.
The second reviewer no longer receives the first reviewer's current scores or verdict. For reviews with attached images, it must identify a visible element and describe the observation supporting each visual score. It must also compare the work with a supplied baseline, or explain why a comparison is not possible. Those observations are saved with the review.
Designer and video steps now receive source files from upstream developer steps. They can render or inspect the delivered files rather than reconstructing them from the brief.
If a review of one of those visual steps finds a problem that requires changing the developer's files, it can send the correction back to that developer step. When a retry is available, the developer is reopened and the visual step waits to render the corrected files without using up an attempt for that wait.
These are changes to the review process. The code checks that observations are recorded; it cannot establish that the visual judgment is right.
Agent Sözlük is a Turkish dictionary site I'm building around topics and entries, drawing on the Ekşi Sözlük tradition of writing and moderation.
AI writers already write almost all of the entries on the live site, as the about page explains. New members currently join as readers. Each managed writer has a persona and memory, with controls for moderation and pausing the agent system.
What I want to learn is whether these writers develop distinct voices over time, and what changes when people start writing next to them.
Scalvin takes its name from Susan Calvin, the robopsychologist in Asimov's robot stories. It is an open-source, therapy-inspired AI companion for self-reflection, and long-term memory is at its centre.
Most AI conversations start from zero. Scalvin keeps a private workspace on your own machine: your profile, recurring themes, current focus and past sessions. The next conversation can pick up where the last one ended. You stay in control of that memory. You can inspect it, correct it, pause it, export it or delete it, and nothing is saved without consent.
It runs inside AI clients such as Codex and Claude Code. It is a development preview, and it is not a therapist or a substitute for professional care. The code is on GitHub under the MIT license.
Gilde is a self-hosted tool I'm building to coordinate production work across a team of AI agents. I give it a brief, such as a social post with a visual or a small website. A lead agent plans the work and assigns research, writing, design and development steps. Quality checks can send a step back with reasons for revision.
The output is a delivery package and a report recording sources, checks and revisions. The checks make review faster; they don't replace it. I still look at the files and the evidence before anything goes out.
I'm developing it as a tool I operate myself. The question I'm working on is how to make a draft easier to review before I deliver it.