---
title: "The Week in Agentic AI Coding: Cerevisor 3.6 Opens Worlds Side by Side and Keeps Agents Moving"
slug: weekly-recap-2026-09-25
date: 2026-09-24
excerpt: Cerevisor 3.6.0 opens several worlds at once, runs two side by side and unsticks agents on five tools. Plus four more releases and three posts.
featured_image: "https://bbtxujdxvidaghmhxkqs.supabase.co/storage/v1/object/public/generated-images/blog-1790293316999-weekly-recap-2026-09-25.webp"
featured_image_alt: "Abstract navy and steel illustration of two flowing streams of circuit-like lanes, with a human hand reaching into the right-hand stream and a single gold thread crossing from one stream into the other."
author: Kinan Hamwi
author_url: https://www.linkedin.com/in/kinanhamwi/
canonical_url: https://cerevisor.com/blog/weekly-recap-2026-09-25
updated_at: 2026-09-24T23:57:45.829602+00:00
---

# The Week in Agentic AI Coding: Cerevisor 3.6 Opens Worlds Side by Side and Keeps Agents Moving

### The week in one glance

- Cerevisor 3.6.0 is the headline: several worlds open at once, two of them side by side in one window, a simpler first run, and agents on Claude Code, Codex, Cursor, Antigravity and Grok Build that get stuck less and say why when they stop.

- It caps a run of five releases since our last recap. Together they let you steer a running agent, hand a whole workflow to a colleague as one file, and now keep two projects running beside each other.

- Three posts on the blog: what owning your own API key protects, how a team takes work back from an agent, and what 3.4 and 3.5 make possible together.

## The theme: an agentic AI coding assistant you can work beside, not just watch

An agentic AI coding assistant plans its own steps and calls tools on its own. The harness around it decides how much of that work you can see, reach and run in parallel. This week’s releases push hard on all three. With 3.6.0, one window can hold several projects, each one a “world” (a saved set of workflows and the folder they work in), and it can show two of them side by side, each running on its own. Agents that used to stall now carry on, or they stop and say why in a plain sentence. Earlier in the fortnight, 3.5.0 let you send a running agent a new instruction, and 3.4.0 made a workflow something you can send to a colleague. The result is an AI coding assistant that behaves more like a colleague you work next to than a job you wait on.

## What’s new in Cerevisor 3.6.0

  [**Cerevisor 3.6.0**](/releases/v3-6-0)
  release, 24 September

Seven improvements in one release, most of them new capabilities rather than fixes. Here is each one in plain terms.

**Several worlds open at once.** The pills in the top-left dock are now your open worlds. Clicking one switches at once, with no save prompt, and the world you leave keeps its unsaved edits, its view and its running work. Each pill shows whether a run in it is going, waiting for you, done or failed, and Ctrl+Tab moves between them. After a restart, the worlds you had open are waiting for you.

**Two worlds side by side.** Choose Open beside, or drag a pill to the right edge of the window, and a second world opens in the right half, fully editable, with its own Save and Close. You can edit, run and chat in both at once, and each save writes only its own file. A Stop on one side now stops only that side’s run.

**A simpler first run.** Cerevisor always starts on Home, and the main button reads “Pick the folder to work in”. Entering a world opens the Workflow Builder chat straight away. The first-project guide is optional. It explains a workflow in everyday language and follows the working agent on the canvas, and it never clicks or opens anything for you.

**Fewer stuck agents.** Claude Code agents keep the ability to ask for approval mid-run. Codex agents allowed to change files can use the internet when web access is on, and the note before a run says exactly what they can reach. Cursor’s Stop works while an agent is still starting, Antigravity picks up a terminal sign-in within 2 minutes, and Grok Build now responds to the Workflow Builder’s Approve and Deny buttons. Agents on Cerevisor’s own engine get four times the token allowance per run. A brief Windows file lock from antivirus or a backup no longer stops a run.

**Your Claude Code settings, your choice.** A new checkbox decides whether workflow agents use your personal skills, plugins and hooks, or only the project folder’s settings. It is on by default, and the note before a run says so. Secrets you assign from the vault now reach Claude Code agents too.

**Decision cards that read what you gave them.** A decision card with no agent before it now judges what you typed at launch. Decisions reach the next agent as one table: question, answer, how sure, who answered. If you turn the optional decision model off mid-run, nothing more is sent to it.

**Dialogs that block, and a calmer canvas.** Every blocking dialog now really blocks what is behind it. While you work in the chat builder or the run window, the rest of the controls fade until you point at them, and an agent asking you something never fades.

## Also shipped since the last recap

  [**Cerevisor 3.5.0**](/releases/v3-5-0)
  release, 23 September

Talk to an agent while it works. A message sent to a running Claude Code agent arrives at its next tool step, so it can change course, and sending a message never grants a permission. Codex agents can stop and ask you a question instead of guessing. Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna are available, and signing in to Claude now uses Claude's own browser login.

  [**Cerevisor 3.4.0**](/releases/v3-4-0)
  release, 21 September

Your work leaves the computer it was made on. A saved workflow opens and runs on another computer as long as its folder travels with it, and Share turns a workflow or a whole world into one file with nothing of your machine in it: no paths, no keys, no run history. Decisions on the canvas show exactly what they were given, and saves are written to the side and swapped in only when complete.

  [**Cerevisor 3.3.0**](/releases/v3-3-0)
  release, 18 September

A better first hour. Connecting an AI model starts by asking what you already pay for and never needs a terminal. A guide walks a new install to a first finished result, quick checks (the small yes or no judgments a run makes) are visible for the first time, and there are five color schemes. The release adds Free, Pro and Founding member plans, and anyone who already held a license is a Founding member and loses nothing.

  [**Cerevisor 3.2.1**](/releases/v3-2-1)
  release, 14 September

A polish release on 3.2.0. The one you will feel first: a Check-by-hand answer can now be changed, and both answers stay in the run's history. The Overview now says when a run was continued, and reviews behave exactly as the guides describe.

## What we published

### AI coding agents this week

  [**Self-hosted AI agents: what bring-your-own-key actually protects**](/blog/harness-self-hosted-ai-agents-byok-architecture)
  harness-security

A reference on where an agent's API key actually sits, what owning it protects, and what it does not.

  [**When the AI coding assistant hands the work back, who on the team can take it?**](/blog/harness-ai-coding-assistant-takeover-gap)
  harness-operator-sustainability

Why taking over from a running agent is a skill worth practising, and a four-step rehearsal any team can run to measure it.

  [**Cerevisor 3.4 and 3.5: send a workflow as one file, then steer its agents mid-run**](/blog/cerevisor-v34-v35-share-and-steer)
  harness-tool-evaluation

What the two releases make possible together: handing a workflow over, then talking to it while it runs.

## Signals to implications for choosing an AI coding assistant

**Signal.** Cerevisor 3.6.0 runs two worlds side by side, each saving only its own file, and a Stop on one side stops only that side's run.

**Implication.** Ask any tool on your shortlist to run two projects at once and stop one of them. If the other one stops too, parallel work is not really parallel. *[Eng Leader]*

Source: [Cerevisor 3.6.0](/releases/v3-6-0)

**Signal.** Owning the API key changes who is billed and who can revoke access. It does not change what a provider keeps or what the software around the model writes to disk.

**Implication.** Before the next rollout, list every destination that receives anything from a machine running agents, optional add-ons included. *[Exec + Eng]*

Source: [Self-hosted AI agents: what bring-your-own-key actually protects](/blog/harness-self-hosted-ai-agents-byok-architecture)

**Signal.** A shared Cerevisor workflow arrives with its rules but not the sender's approval of them, so the receiver's first run shows the normal pre-run summary.

**Implication.** When work moves between people, consent should stay with the person. Check whether any tool on the shortlist carries approvals from one machine to another. *[Founder]*

Source: [Cerevisor 3.4 and 3.5: send a workflow as one file, then steer its agents mid-run](/blog/cerevisor-v34-v35-share-and-steer)

## The contrarian take on the agentic AI coding assistant

Most comparisons of an agentic AI coding assistant stop at the model. This fortnight says the room around the model matters as much. [3.6.0](/releases/v3-6-0) lets two projects run side by side, [3.5.0](/releases/v3-5-0) lets you redirect an agent mid-task, and [3.4](/blog/cerevisor-v34-v35-share-and-steer) lets the work travel to a colleague. [The takeover post](/blog/harness-ai-coding-assistant-takeover-gap) adds the human half: those controls pay off most for a team that has practised using them. The next move: pick one real workflow this week, open it beside another, steer it once mid-run, and hand it to a teammate.

## Next week

If this recap was useful, the newsletter delivers it straight to your inbox every week. [Subscribe here.](/#newsletter)