--- title: "The Week in Local AI Agents: Cerevisor 3.8 Gives Every Agent a Lasting Chat" slug: weekly-recap-2026-10-11 date: 2026-10-11 excerpt: Cerevisor 3.8.0 gives each agent a lasting chat, and 3.8.1 repairs the Mac build and makes that chat one view. Plus two posts from the week. featured_image: "https://bbtxujdxvidaghmhxkqs.supabase.co/storage/v1/object/public/generated-images/blog-1791738670860-weekly-recap-2026-10-11.webp" featured_image_alt: Flat navy and steel blue illustration of a wide road, a straight track and a winding footpath crossing an open plain and meeting at one gate with a small gold light above it. author: Kinan Hamwi author_url: https://www.linkedin.com/in/kinanhamwi/ canonical_url: https://cerevisor.com/blog/weekly-recap-2026-10-11 updated_at: 2026-10-12T10:01:39.960488+00:00 --- # The Week in Local AI Agents: Cerevisor 3.8 Gives Every Agent a Lasting Chat *By the founder of Cerevisor. Release details checked on 12 October 2026 against the Cerevisor 3.8.0 and 3.8.1 release notes.* ### The week in one glance - Cerevisor 3.8.0 is the headline. Each agent card now has one conversation that survives a restart, a Stop and later runs, and every message sent goes through the same checks a run does. - Cerevisor 3.8.1 followed on 11 October. It repairs the Mac build, above all on Intel Macs, makes an agent's chat one view with a time on every entry, and lets a step that hits its usage limit continue on one of your subscription apps. - Two posts: what an agent's stated confidence is worth, and the reasoning behind the 3.8 chat and help assistant. ## The theme: local AI agents, with one set of rules for every way in By local AI agents I mean agents that are run from this computer, under rules set on this computer. This week was about making the convenient way in follow the same rules as the formal one. In 3.8 a message to an agent is held to the rules of a run: the agent answers with the tools, permissions and instructions on its card, and each reply is listed in the System Log as a Chat turn. Claude Code and Codex agents can now use connected tools, with each call made by Cerevisor through the same checks as any other tool call. And the post I wrote on Wednesday reaches the same place from the research side: an agent saying it is sure is no reason to skip a second opinion. Then 3.8.1 made the new chat simpler to read and repaired the Mac build. ## What’s new in Cerevisor 3.8.0: a lasting chat, held to the rules of a run [**Cerevisor 3.8.0**](/releases/v3-8-0) release, 8 October 3.8 gives every agent a conversation that lasts. Here is each part in plain terms. **One conversation per agent.** Each agent card has one conversation, kept on this computer. It is still there after a restart, after a Stop, and after later runs of the workflow. It works with an agent on every kind of provider, including Claude Code, Codex, Cursor, Antigravity and Grok Build. A message to an agent that has finished now continues the conversation. **Easier to type into.** Replies stream as formatted text with a card for each tool call. The box grows while typing and is never locked, and any file can be pasted, dropped or picked. One status line says what the agent can do right now. **Held to the rules of a run.** The agent answers with the tools, permissions and instructions on its card. What agent chats cost is shown in Analytics, under Total Cost. **The AI already on the computer.** On a computer that has never had a provider in Cerevisor, the first launch looks for Claude Code, Codex, Grok Build and Ollama. Each one that is installed and signed in is connected without a click, and one notice names it. This happens once per computer. **Connected tools.** Add MCP server is now one box that takes a URL, a config or a command. A token typed there is kept in the operating system’s secure storage and is used only at the address it was saved for. Claude Code and Codex agents can now use connected tools. **Smaller changes.** The run report opens when a run started in the window ends. Import shared workflow adds a shared file’s workflows to the open world. The help assistant, Need a hand?, now gets a short description of the screen with each question, so it can explain a run or list a workflow’s problems without being told where to look. Every change, with the known limits, is in the [3.8.0 release notes](/releases/v3-8-0). ## What’s new in Cerevisor 3.8.1: Mac repairs and one chat view [**Cerevisor 3.8.1**](/download) release, 11 October 3.8.1 repairs the Mac build and tidies what 3.8.0 introduced. It is the version on the download page now. **Mac fixes.** On an Intel Mac, 3.8.0 could not sign in to Claude or Google. 3.8.1 repairs that, so anyone on an Intel Mac should download 3.8.1. Command+Q now quits. macOS asks for folder permission less often, and each notice says why. The app starts quicker. **Updating on a Mac is by download for now.** Download the new version from [cerevisor.com/download](/download) and replace the app in Applications. Workflows and settings are kept. On Windows and Linux the built-in updater brings 3.8.1 as usual. **A backup when a provider hits its limit.** In Provider Hub, under If a provider hits its limit, every enabled connection is now listed, so Claude Code, Codex, Cursor, Antigravity and Grok Build can be the backup. The step runs again on the backup app, under that app’s own permissions, and an agent moves once per run at most. **An agent’s chat is one view.** The Activity and Chat tabs are gone. While the agent works in a run, its steps stream in under the conversation, and when the run ends they fold under that run’s line, with Show the steps to open them again. Every entry has a time. **A schedule runs its own workflow.** In a world of several workflows, a schedule runs the workflow it was made for and no other. **Smaller changes.** The Windows installer is one card: Install, then Open Cerevisor. The Board tab and the canvas conversation panel have a Pinned notes section. The download is smaller. ## What I published ### AI coding agents this week [**AI agent evaluation: what an agent's stated confidence is worth**](/blog/harness-ai-agent-evaluation-stated-confidence) harness-operator-sustainability How much weight should an agent's own "I am sure" carry? The post sets two weeks of my own runs beside three studies. Its conclusion: a low number is a lead worth following, and a high number is no reason to look away. [**Cerevisor 3.8: a proper chat for local AI agents, and help that sees your screen**](/blog/cerevisor-v38-local-ai-agents-lasting-conversation) harness-tool-evaluation The reasoning behind the release: how the new agent chat works, why a chat message is held to the rules of a run, and what the help assistant can now see and explain. ## AI agent evaluation and MCP server authentication: three signals **Signal.** One study the post quotes found that "longer explanations increased user confidence, even when the extra length did not improve answer accuracy." A workshop paper it presents at lighter weight reports that some agents that succeed only 22% of the time predict 77% success. **Implication.** Let stated confidence add checks and never remove one. Whatever decides that work is good enough should come from somewhere other than its author. *[Eng Leader]* Source: [AI agent evaluation: what an agent's stated confidence is worth](/blog/harness-ai-agent-evaluation-stated-confidence) **Signal.** In Cerevisor 3.8, the tool calls of a chat message go through the same permission checks, approval setting and tool hooks as a run, and each reply is listed in the System Log as a Chat turn. **Implication.** When comparing agent tools, ask whether a typed message, a scheduled run and a run started by hand all follow the same rules and leave the same record. *[Exec + Eng]* Source: [Cerevisor 3.8: a proper chat for local AI agents, and help that sees your screen](/blog/cerevisor-v38-local-ai-agents-lasting-conversation) **Signal.** A remote tool in 3.8 has an Authentication setting: sign in with the browser, a bearer token, or none. A token is kept in the operating system's secure storage and is used only at the address it was saved for. GitHub's card asks for a token and says where to make one. **Implication.** Before wiring a remote tool into a workflow that runs unattended, connect it once at the desk and run it by hand. *[Founder]* Source: [Cerevisor 3.8.0 release notes](/releases/v3-8-0) ## The contrarian take on local AI agents: a better chat is not a reason to chat more It would be easy to read a lasting, comfortable chat as an invitation to do the work in the chat. I read it the other way. The chat is where I ask an agent why it made a choice and steer what it does next, and [the 3.8 post](/blog/cerevisor-v38-local-ai-agents-lasting-conversation) explains why each message is held to the rules of a run. The workflow is still where work that repeats belongs. And a confident reply in a chat deserves the same second look as [a confident result in a run](/blog/harness-ai-agent-evaluation-stated-confidence). The next move is small. After updating, talk to one agent for a day, and put anything you asked for twice into the workflow. ## Next week If this recap was useful, the newsletter delivers it straight to your inbox every week. [Subscribe here.](/#newsletter)