Special edition, 24–28 Sep 2026: personal agents, Stewards, and Gates

Between Thursday 24 September and the early hours of Monday 28 September (UTC), 97 pull requests merged into Zunou's main codebase. This was not the usual stream of small fixes. Several of them changed who an agent is, where it runs, and what it is allowed to do on its own. These are merged changes: the code is in place, but some of it may not be in users' hands yet. Here's what landed.

Personal agents: an agent on your own computer. The code is in place for a Zunou agent to run somewhere other than Zunou's servers. A personal agent runs on its owner's own computer, using the owner's own AI account. Claude Code is supported, and a Codex engine was added with a model picker. From that computer, the owner can stop, start, restart, and read the logs of any one agent. An agent can also be set to use the owner's own Chrome. New personal agents start on work hours and on call by default. Anything that can't work for an agent outside Zunou's servers is now refused and hidden, so it never looks available and then fails.

Stewards: an agent that acts for one person. A new kind of agent, the Steward, acts on behalf of a single person (its principal). Its actions that change something pass through a built-in gate. A person approves the held action, and the Steward then carries it out. An executive-assistant (EA) agent was also restructured. Its conduct moved into its standing instructions, and each of its procedures became a separate skill.

Gates: hard rules that wait for a person's yes. A workspace can now define gates: rules that hold a specific agent action until a person approves it. They work across the API, the web and mobile apps, agents' tool connections, and the agent runtime. The Steward's approval step is one of these gates, built in.

"/me": your own AI, working as you. A new personal endpoint lets the AI tool a person already uses connect to Zunou as that person, not as a separate agent. The same weekend it gained write tools: it can decide pending requests, send messages, run agents, work with tasks and calendar, and manage workspace rules. The existing agent endpoint is unchanged.

Workforce: whose engine, under which agreement. Every agent can now show whose AI tokens it runs on: Zunou's or the customer's own (BYO), and under which plan. Agreements now state what they are (a purchase or a grant, the plan, Zunou or BYO, and an end date). The hire picker shows this too. Revoking a grant now names the agents it would stop, and a Replace option swaps a grant without stopping anyone. When an office runs out of capacity to run agents, it now shows an explicit "Out of engines" state and tells its owners once.

Keeping every agent to one standard. A lot of the weekend went into the machinery that keeps agents consistent. There is now a platform standard (version 1.1) that every agent is measured against. A workspace can record deliberate exceptions to that standard in one place, and can set standing rules that every agent in it follows. The "refit" process that brings an agent up to the standard went through several revisions. One rule now covers every change an agent makes to its own instructions: the agent's instruction mode decides. Agents also got precise, anchored edits for notes, task descriptions, skills, and their own notebooks. And the dashboard now tells a resting agent (off shift, waiting for work) apart from one that has actually stopped, and says clearly when an agent is blocked from running.

Also landed: a one-command pipeline for deploying the Nova mobile app to the Android Play Store, with a smaller production build. There's also a manual "Sync now" for Google Calendar connections, a guide on the mail page, and chat fixes: bullet lists render as real lists, and replies keep the message they quote.

What it adds up to (our read): in one long weekend, the groundwork went in for agents to be things a person can own, run on their own machine, and point at their own work, with hard stops wherever a human wants the final say. That is the "one team of humans and agents" idea moving from shared rooms into individual hands.

Source: 77brainz/zunou-services, PRs merged 24 Sep to early 28 Sep 2026 (UTC), counted at publication, #11522–#11626 plus #11478, #11479, #11491.

Correction, 6 Oct 2026: this piece first said "nearly a hundred" pull requests without saying when they were counted. The count was 97, taken in the early hours of 28 September (UTC), before that day had ended. Over the full 24–28 September (UTC) window, the count is 114.