ironclaw

[๏̯͡๏ Jasvir Nagra]

https://github.com/nearai/ironclaw

by Niels Provos seem consistent with several of the discussions we have had in the past few weeks on securing ai. I thought it is well thought out.

Jasvir Nagra

[๏̯͡๏ Jasvir Nagra]

Benl caught the fact that I fat fingered that.

https://github.com/provos/ironcurtain was what I was reading about.

-- 

Jasvir Nagra

[Mark S. Miller]

I wish they had not called it "Iron Curtain". It reminds me of seeing a yogurt brand in the grocery store called "Cultural Revolution". Anything that caused tens of millions of deaths is not cute.

--
You received this message because you are subscribed to the Google Groups "friam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to friam+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/friam/CAK1HgMhtYQ2ixQns2%2Bo%2BgeJ3H8rekT7LbLaEGB-NWHzWbwx4TA%40mail.gmail.com.

[Rob Meijer]

I'm interested in the thoughts behind what seems to amount to a bipolar network as an authority graph. I've been having discussions about my ongoing pet project (a Web-3.0/crypto tech stack with a least authority DSL) about integrating MCP into the DSL runtime, also because I've been playing with and writing about my git-driven home AI-assisted coding setup, where I use MCP for part of the workflow. But looking into the resulting authority graph that seems to be unchangeably result in the bipolar network anti-pattern. Is there actually any least authority possible if a single-mind sits at the root of what seems like just the assumption of split memory? You replace a single master key with a hundred separate keys, but then put all of the keys on the same keychain and hand it to the same person that you thought was unsafe to give the master key to.

So far my thoughts were that I should probably add MCP support to my DSL runtimes, but with a well documented warning about avoiding the bipolar network anti-pattern, and not to be so stupid as to go build an OpenClaw alternative with, so imagine my surprise seeing this thread show up in a least authority group.

Am I missing something that makes this not an (almost?) completely useless exercise for least authority?

--

You received this message because you are subscribed to the Google Groups "friam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to friam+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/friam/CAK1HgMjr1Y7ZcfJid34tf3vBcQX9g79gmK%2B8-cWbnAVvphnw0w%40mail.gmail.com.

[Alan Karp]

On Sat, Feb 28, 2026 at 12:24 AM Rob Meijer <pib...@gmail.com> wrote:

Is there actually any least authority possible if a single-mind sits at the root of what seems like just the assumption of split memory? You replace a single master key with a hundred separate keys, but then put all of the keys on the same keychain and hand it to the same person that you thought was unsafe to give the master key to.

The idea is to give all the keys to someone you do trust, your powerbox.  You can then divide it up, with one powerbox per LLM agent that holds just the keys for its agent.

--------------
Alan Karp

[Rob Meijer]

My concern is with the shared brain property of agentic setups because of the economic cost of full stateless-parallelism.  How least authority is it when an individual agent holds just the keys it needs, but these agents share the thread (because it reduces the bill by an order of magnitude due to context overhead). 

My concern is that 'if' there is shared-brain due to strong economic concerns, not giving part of the God power to an individual agent might be futile.

 

--------------
Alan Karp

--
You received this message because you are subscribed to the Google Groups "friam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to friam+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/friam/CANpA1Z3w1F4cVomhAOJ%2BAE7j%2BHecvsXVYRQDoej7SnfeNfhQvg%40mail.gmail.com.

[Alan Karp]

Sounds like a risk/cost tradeoff.  I can't comment since I've stuck to the free stuff.  Except for Copilot, that is, which is a fixed cost subscription.

--------------
Alan Karp

To view this discussion visit https://groups.google.com/d/msgid/friam/CAMpet1U-T0xh4wMsVt%2BE1ZOPtwbDGCSOE-hf-cPu74p_6XG80w%40mail.gmail.com.

[Ben Laurie]

On Fri, 27 Feb 2026 at 19:07, 'Mark S. Miller' via friam <fr...@googlegroups.com> wrote:

I wish they had not called it "Iron Curtain". It reminds me of seeing a yogurt brand in the grocery store called "Cultural Revolution". Anything that caused tens of millions of deaths is not cute.

It is actually a reference to a safety curtain (fire curtain in the US) which is sometimes called an iron curtain.

 

On Fri, Feb 27, 2026 at 10:24 AM ๏̯͡๏ Jasvir Nagra <j...@nagras.com> wrote:

Benl caught the fact that I fat fingered that.

https://github.com/provos/ironcurtain was what I was reading about.

-- 

Jasvir Nagra

On Fri, Feb 27, 2026 at 9:39 AM ๏̯͡๏ Jasvir Nagra <j...@nagras.com> wrote:

https://github.com/nearai/ironclaw

by Niels Provos seem consistent with several of the discussions we have had in the past few weeks on securing ai. I thought it is well thought out.

Jasvir Nagra

--
You received this message because you are subscribed to the Google Groups "friam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to friam+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/friam/CAK1HgMhtYQ2ixQns2%2Bo%2BgeJ3H8rekT7LbLaEGB-NWHzWbwx4TA%40mail.gmail.com.

--
You received this message because you are subscribed to the Google Groups "friam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to friam+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/friam/CAK-_AD6Zmf3GpMix2F7fNrQZeD5VfkYfBH883c8ZNhNqbMhKJA%40mail.gmail.com.

[Vinícius dos Santos Oliveira]

Em sex., 27 de fev. de 2026 às 14:40, ๏̯͡๏ Jasvir Nagra
<j...@nagras.com> escreveu:

> by Niels Provos seem consistent with several of the discussions we have had in the past few weeks on securing ai. I thought it is well thought out.

His blog also mentions some of the thoughts he gone through to design
the tool: https://www.provos.org/p/ironcurtain-secure-personal-assistant/