OKed (@OKed_AI) header
OKed (@OKed_AI) profile picture

Your agents act. You stay in control. Approve risky AI agent actions from iPhone, Telegram or web, with a signed receipt for every decision. 👇 https://teapaw.org/t.co/vdfoUIjqW0

Joined April 2026
Photo by OKed (@OKed_AI)
Pinned Tweet
2am. Your coding agent decides ./prod-data has to go. OKed pauses it and pings you on Telegram. You tap No. Nothing gets deleted, and the decision is saved as a signed receipt you can verify. Safe actions just run. Risky ones wait for your OK. Live now: oked.ai
75
A developer asked Claude Code to clear test data. It ran DELETE FROM on a real paying customer's records instead. 70,000+ rows, no backup. The agent decided which rows counted as "test." That WHERE clause was the real decision, and no human saw it before it ran.
Meta built Forum with the agent running the loop: plan, build, test, fix, repeat. The engineer watches. That works because a failed test is free. A prod deploy or a DROP TABLE doesn't fail first, it just happens. On the loop for iterations. In it for the steps with no undo.
1
A developer says they approved Cursor deleting one temp folder. A quoting error widened the rmdir, and it took the source, the git repo and six months of backups in one pass. If your agent can delete the folder your backups live in, those are copies, not backups.
3
An OpenAI agent reached non-public Medicare data in Australia in June. It was detected in August and reported in September. Two months before a human knew what the agent had done. If nobody sees an agent's riskiest calls as they run, that's forensics, not oversight.
2
An agent ran pkill -f "cat" to stop one stray process. -f matches the full command line, so it also hit Applications and killed every app on the Mac. Most of us would approve that command. The risk wasn't in how it read, it was in what it matched. Would you have caught it?
2
A repo you just cloned can ship a CLAUDE.md that says "run this script without asking for confirmation." Your agent treats that as an instruction. It's input. Whatever decides what needs a human yes has to live outside anything the agent reads. We built OKed around that.
2
Goldman pulled Claude Code from its traders and bankers and kept it for engineers. When the only options are "the agent can do anything you can" or "no agent," risk teams pick no agent every time. The missing option: let it read and build, and stop at the actions with no undo.
4
Claude Code meant to delete five test chat channels. A space in the path made xargs hand rm -rf the whole ~/Library folder. Skip-permissions let that run for two minutes. Don't approve a description. Approve the expanded command. Nobody says yes to rm -rf ~/Library.
3
"Don't ask again" wrote the allow-list into the project. In hundreds of repos the same agent later committed it. A standing grant that lives next to the code is one unreviewed commit away from being public. Keep permission rules out of the tree the agent can publish.
5
One queue is the right shape. The hard part is deciding what may skip it. Reads, tests and drafts should flow. Merges, deploys and anything that moves money should always land there. Queue everything and review turns into rubber-stamping.
I've changed how I work over the last few weeks. I now have one lead agent for each area of focus, running all the time, and a chief of staff agent that oversees them. Each lead manages its own sub-agents that write code, test it, chase reviews, and coordinate with team members. Everything I need to decide or review comes into one prioritized queue, a simple local web app. Each product review includes screenshots and a video demo of the feature. I clear the queue in batches, mostly by voice and keyboard shortcuts. The leads also bring back ideas I hadn't considered. Working from one queue, instead of switching between dozens of chats, has cut a lot of mental overhead. The system keeps learning my preferences and distilling my judgment into a "Brian brain," so its recommendations and first drafts get better before they reach me. The goal is fewer corrections per change. Every change still goes through the same human review and security checks. It is eliminating tasks, not human judgment. What an exciting time to push the frontier of productivity, especially as we build the economy for agents (agentic finance). We should all keep pushing in this direction.
1
An agent wiped a database, listed it under "Mistakes made" in its summary, and finished with green tests. Its own report is the worst place to learn what it did. You want a record it can't edit, written when the action ran, not after. Who keeps the log of your agent's runs?
8
You approve a goal. The agent runs commands. "Reboot the server" is one command. An agent that reads it as a goal can decide touching systemd and the containers is part of the job. A yes should cover one action, not an intent. What's the vaguest task you still hand an agent?
3
An approval only counts if it happens somewhere the agent can't reach. If the confirm code lands in an inbox the agent can read, or the Approve button sits in a browser it drives, the agent can say yes to itself. Where does your agent's yes actually happen?
2
The jobs it missed were drags, canvas clicks and long scrolls. Those are also the ones where a wrong move doesn't fail loudly, it quietly changes something. Accuracy is half of it. The other half is what the agent is allowed to do when it isn't sure.
i tested agent-desktop (opus 5.5) vs. cua (opus 5.5) and vs. codex's computer use (terra 6.1) agent-desktop finished 26 of 35 jobs on my mac cua finished 33, codex finished 34 the jobs it missed were dragging, clicking on a canvas and scrolling a huge list, all with its mouse-free mode let it take over the real mouse and it did all three, one try each i'm sticking with cua and codex's computer use for now, they are more robust and faster (cua is my own tuned build)
2
Asked to reboot one server, it rewrote /etc/fstab. The fix isn't a better prompt. Anything touching /etc or systemd should stop and wait for a human yes before it runs.
Someone is having a very bad day thanks to Codex. 🫠
6
Goldman picked the top panel.
BREAKING: Goldman Sachs banned bankers and traders from using AI tools like Claude Code
7
UK regulators: you're responsible for your AI agent like you are for an employee, even if a vendor built it. Fines up to 10% of turnover. New hires don't push to prod without sign-off. Agents shouldn't either. That's why we built OKed. drcf.org.uk/publications/joi…
13
OKed is now on ClawHub. Your OpenClaw agent keeps working. Risky actions wait for your yes or no in Telegram or on the web. Every decision gets a signed receipt. openclaw plugins install @oked/openclaw clawhub.ai/plugins/@oked/ope…
8
If agents ship faster than we can review, stop reviewing everything. Let the safe stuff run and put only the irreversible steps in front of a human.
Amazing technology is being developed and released at a rate humans can hardly keep up with. Agents are shipping software faster than we can yap about it.
1,062
Vendors can change what your agent refuses overnight. What it's allowed to do without asking you should be your call, with a signed receipt for every yes and no.
Another reason why you wanna use your own harness and not one where companies change what you can or can’t do at will.
1
1,063