Tech

Agents are fluent. Fluency is not the same as safe. If you are rolling AI agents into a real org, you already know the demo version of this problem. The model sounds careful. The slide says “guardrails.” Someone writes “do not touch production” in a prompt, and the room relaxes. That sentence is a sticky note. It is not a control. When the agent is in a hurry, or the chat got cut short, or it found a clever way to spell the same dangerous action, the sticky note is already gone. ...

The search found the line it was looking for. That was the annoying part. I run a handful of small safety checks around my coding agents, and one has a simple job: if the agent that wrote a pull request tries to merge it, stop it. A pull request is a proposed change waiting for approval, and the point is that the author does not get to be the approver. Under a heavy test run, that check started getting the wrong answer. It found the agent’s own work right there in the session’s commits, then concluded the agent had written nothing, which is the one answer that lets the merge through. ...

The build was green. That was the annoying part. Somewhere under a film render pipeline I care about, a case that checks how a scene render names its output broke on the default branch and stayed broken for about six hours, until someone ran it by hand. CI never flagged it, because the file that held that case had sat off the continuous-integration checks for a long time. The dashboard was not lying about the jobs it ran. It was silent about the jobs nobody had asked it to run. ...

I used to want one agent that could do everything. Write the PR, tidy the calendar, answer Slack, review the diff, remind me about lacrosse pickup — one conversational surface, one “brain,” one place to yell when something was wrong. It felt efficient until the failures started rhyming. Coding work leaned on personal-assistant shortcuts. Calendar noise interrupted engineering focus. A gateway that looked like a helpful front door became a noisy hallway where every request sounded the same, and “the agent is up” — meaning the process is running, the light is on — stopped meaning either job was actually done. ...

I have shipped test suites that were green for the life of a repo and still left a hole on the LAN. By green I mean the dashboard color everyone treats as “safe to ship” — the check passed, the badge is cheerful, nobody is paging. CI (continuous integration) is the automated check-runner that usually paints that color: on every push or pull request it runs a suite of tests and jobs, then reports pass or fail. Green CI is valuable when it asked the hard questions. It is dangerous when it only asked the easy ones, or never ran the suite that looked like coverage. ...

Chat memory is convenient until you change apps. I run a Grok Bot roster every week — Chief of Staff, specialists, digests, drafts — and I also point other agent platforms at the same durable facts about how I work. If those facts only live inside one vendor’s chat memory, I am not building a system. I am renting amnesia with a nice UI. So I keep a second brain I own. Mine is called MainVault: a markdown vault, synced with git, treated as shared memory across agents. I already wrote the high-level “owned notebook both stacks get” frame when piloting Muse beside Grok Bot. This post is the mechanism: how that notebook works day to day. No private repo links. The point is the pattern. ...

I am running two personal agents at once right now: Grok Bot and Meta’s new Muse. That is not a brand loyalty bit. It is a bake-off with an honest mismatch of maturity: I am piloting Muse next to a mature multi-bot Grok Bot desk I already run. This is not a feature-parity scorecard. Muse is new and simple. Grok Bot is a roster I have already wired. I want to feel where each one wins on its own terms before I pretend there is a single winner. ...

Companion to GitHub tokens for agent fleets and the agent production system series. Words I use below Agent — a program that writes and lands code for you, in a chat session Pull request (PR) — a proposed change waiting for review; on GitHub it has its own page, with a button to accept it Merge — accepting that proposed change so it becomes part of the main project GitHub App — a robot you install on a GitHub project so it can click buttons as itself — not as you; it can approve, and it can merge; people treat this robot as the “real” setup, and it is not — it is extra (this post does not tell you how to install one; if you later want that robot, read the token post) Lock — a program that says no when a rule is broken; the agent cannot talk it into a yes; I also call this the landing floor Stamp — a review that applies to one exact version of the code; “looks good” in chat is not a stamp Automatic tests — checks that run by themselves when you propose a change; green means they passed Those seven words are the whole toolkit. Now the point. ...

Companion to The last click is still yours and the agent production system series. Words I use below Agent — a program that writes code for you in a chat session Pull request (PR) — a proposed change waiting for review; on GitHub it has its own page, with a button to accept it Merge — accepting that proposed change so it becomes part of the main project Personal login — you, signed in as yourself; when you type commands, GitHub sees your name GitHub App — a robot you install on a GitHub project so it can click buttons as itself — not as you; it can approve, and it can merge Short-lived token — a temporary key the robot gets to prove it may click; it dies in about an hour, and after that it is junk PAT (personal access token) — a long-lived key tied to you; if it leaks into a chat log, it keeps working until someone turns it off Broker — a small program that hands out a short-lived token when the rules say yes; the agent asks, the broker decides, and the agent does not invent a key Automatic tests — programs that run by themselves when you propose a change; green means they passed Those nine words are the toolkit. Now the point. ...

This is part 6 of the Agent production system series. Previous: Claim safety. Later: GitHub tokens and the last click. Words I use below Agent. A program that writes and lands code for you, in a chat session. Automate. Let the agent do the step without you. Work path. The steps from idea to shipped. I also call this the funnel. The middle is plans, code, and tests. The edges are the start and the finish — where someone can get hurt. ...