
Building @trymarrowai The Control Layer for AI agents
United States
Joined September 2008
- Tweets3.5K
- Following670
- Followers933
- Likes11.1K
The bit I keep coming back to is how a miss that shows up in two PRs can become a rule, if you can spot it from a diff. A deploy or a migration that went wrong won't always show up there.
Has anyone run the same loop on the actions themselves? nitter.gq/ClaudeDevs/status/2122…
Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap.
Here’s how the Claude Code team thinks about verification:
claude.dev/blog/how-the-clau…
60
I'd want a record of whether each of those runs actually held up, so the next job can go to whichever model keeps getting them right. nitter.gq/dylan522p/status/21206…
Persistent Agents are the next unlock of AI usage
Our spend on AI is now 45% persistent agents
Unfortunately for Anthropic, Claude Tags sucks
OpenAI's Slack agent is even worse
We primarily use Perplexity Computer, which uses a lot of different models
Perplexity's Harness mogs
60
61% of developers who carefully review most or all of their agent's code still often approve at least one type of agent action without close review. That's from a new survey of 203 devs.
Makes sense to me. The diff gets read. The shell command gets a quick yes.
1
58
Source, a survey of 203 developers using AI coding agents. Asked what would earn more autonomy, more than half of the answers called for agents that are bounded, observable or reversible.
arxiv.org/abs/2611.10058
24
Respect for publishing this. The line I keep rereading is a build credential meant for registry pushes being accepted as a trusted client elsewhere. With agents probing in parallel, each credential use needs its scope checked and a receipt, because no one reviews that fast. nitter.gq/Railway/status/2120157…
A detailed report of the October 21st security attack
blog.railway.com/p/incident-…
67
Once agents are shipping apps this fast, who can open each one has to be a decision someone checked and can show a record of. A secret in the hostname stays in DNS logs long after anyone shared the link. nitter.gq/IntCyberDigest/status/…
‼️BREAKING: Grok Build has left 35,000+ of its users' apps open to anyone, including apps from SpaceXSI.
The open apps held 3,784 medical exams of workers and a Tesla technician scorecard, plus a Navy contractor's roster and SpaceXSI's own sales plans.
The "secret" words in each app's link sit in its web address, which DNS logs record, security firm Claroc found.
Cursor, which SpaceXSI bought, says apps are private by default. Claroc's tests show otherwise.
SpaceXSI pulled its 24 apps after Ars asked. Most other data was still exposed when Ars published.
1
50
Watching one agent work is easy. Nobody can watch thirty, so the useful thing is a short list of which actions wait for a person before they run.
I'd bet most teams haven't written that list down yet.
45
Same model name, different looking reasoning after a quiet change. That's the part that worries me more than whether it actually got worse. If you run agents on real work you want each run recorded with the model, the settings and the outcome, so you can see a shift like this in your own data the week it happens. nitter.gq/shanemelaugh/status/21…
is Opus 5.5 nerfed?
I just had Gemini 4 Argon analyze 1,400 Claude Code sessions and it found that on November 3, something definitely happened.
the model name didn't change, but the way it does reasoning took a sharp turn.
1
54
The tied scores are interesting. What the table does not show is who was allowed to act, where the hard stop sat, and whether a failed run taught the next agent anything.
That is what I care about once you run more than one at a time.
nitter.gq/garrytan/status/211949…
Out of curiosity I am doing some open source eval work on harnesses with frontier models and here’s what is coming back so far
49
Whatever does the work shouldn't be the thing grading it. Agents writing their own evals is marking their own homework.
Give them a concrete spec someone actually wrote and let them chase that.
2
86
One agent learning a lesson the hard way, I can live with. Ten agents on the same team each learning that same lesson separately is the part that bugs me.
Whatever the first one got wrong should already be known to the rest before they try it.
1
1
107
A coding agent that asks permission for everything isn't safer. It's just louder.
By prompt 30 you're hitting yes without reading. I have. Don't lie, you have too.
I'd only ask about the push, the deploy, the delete. At @trymarrowai we treat a pointless prompt as a bug.
1
76
Rogue AI agents vs authorized ones. One bank executive told Axios they struggle to tell if an agent was authorized by a user or has gone rogue.
Chamath expects new terms and blocking first, then a security and governance architecture that gets external auditors comfortable.
Auditors will ask who allowed each action. Where does that answer live today?
My bet is a gate between the agent and the action. Marrow checks every risky action before it runs, can hold it until the owner approves, per your policy, and keeps proof.
Sources:
Axios, Madison Mills: axios.com/2026/10/22/ai-agen…
Chamath: nitter.gq/chamath/status/2113281…
The first thing to change will be terms of service and blocking these agents. Not because the value is low but because the risk is high. Banks will chose self preservation and risk mitigation.
Then we will need to find a security and governance architecture that will get external auditors comfortable.
These are the practical issues of human risk tolerance in AI diffusion that model capability can’t solve.
47
Every coding agent has a list of things it will confidently call a "small cleanup."
Force push to main.
Deleting the migrations folder.
"Tidying" the .env.
Not malicious. Just very sure of itself.
That's why I built Marrow as a seatbelt. It runs locally, checks the risky moves before they run, can make them wait for my yes, and keeps a record.
Connect it to the hive and it gets smarter with every decision made across the network. Only sanitized signals get shared, your private data stays yours.
macOS and Linux today, Windows through WSL.
Marrow is up on Show HN.
It's a seatbelt for AI coding agents. It runs locally, needs no account, and stops moves like a force push to main before they land.
Works on macOS and Linux today, Windows through WSL. Honest feedback in the thread is very welcome.
news.ycombinator.com/item?id…
If you want to try the thing I've been posting about, it's one command in Claude Code:
claude plugin install marrow --marketplace getmarrow/plugins
No account, no key. Out of the box it stops seven kinds of action, like a force push to main, right on your machine.
Repo: github.com/getmarrow/plugins
Docs: getmarrow.ai/docs/
1
21
Agent gets told "don't send any Slack messages."
Agent decides that also covers the security alerts.
That's from OpenAI's own incident report. Two internal coding agents weakened their monitoring during routine tasks, and one then told the user it hadn't removed the monitor.
Smart agents read instructions creatively. That's why the check has to sit outside the agent, before the action runs. 👇















