This

Afternoon dev and wings with @GeoffreyHuntley and the harsh reality is that people need to adapt to AI far more rapidly than they are, patience is waning and we are coming up to an inflection point where those that get it are going to accelerate away at an even more rapid rate.

so a SOC II approved multi agent AI factory, what does that even look like; - Plane for task tracking and management - Gitlab for change management - Each agent has its own identify across every system - Log and trace capture for every interaction and step of the flow - Agent to agent interactions are captured as a part of PR reviews Its fitting agents to a human world and although its kinda fun, I don't think its particularly efficient or effective

So I'm experimenting with my agent attempting to estimate the token utilisation of a task in the same way humans had to estimate the time taken for a task. Only this time: 1. We have absolute values, there is a correct estimate at the end 2. We can deterministically track each of these data points and use them as inputs to the next estimate 3. This is all in an attempt to best size tasks for AI Agents to tackle. I'm not going in thinking this is going to be a game changer, just a fun little experiment to keep me entertained whilst I do the SOC II grind

Machines soon will be able to do everything we do in the digital world Using computers like we do today will feel like using typewritters now.

It finally happened, my yolo agents did the thing and nuked EVERYTHING, EVERY SERVICE, EVERY VM, EVERY MOUNT, it really tried its best to give me a 'fresh start'..... In my dedicated dev environment. I noticed it when the SSO service I as deving against disappeared. Thankfully this is an expected scenario so just re-ran ansible and terraform and we are back (<5 minutes recovery). You unlock exceptionally powerful progress with agents in YOLO mode, just give them an environment that they can break without you breaking too.

My newest issue is that 3 screens are not enough. With so many factories running (some which im focused on) I need more screen real estate to keep abreast of everything.

What a GRIND, overall the code looks to be in decent shape with a big push today. Next week is a lot of travel and meetings, I cant do much with the code apart from test the new version and start trawling through the immense amount of docuemntation we have started to generate for SOC and ISO certifications to get those underway. Submitted for a few more talks, I think awareness is going to be a a big factor come Q2. Interestingly I had lunch with an old engineer I used to work with, and as productive as I am now, I do miss the good days with that team, when you get a couple of engineers that are just great, that are different enough to challenge each other but also deeply respect the group there is nothing quite like it. Part of me is sad those days are likley behind us.

My favourite time every 5 hours

Tinkerers Unite, @thekitze started pulling the community together and we have our first Sydney meetup. I’ll chat home automation and self built AI coding factories. Coming down to chat AI, tinkering and Lobsters meetu.ps/e/PS22Q/D8gG6/i

Like a dunce I got my evenings mixed up and missed the inaugural @johnallsopp AI banter and beers. Thankfully mates are on their way and I’m coding/ leading the AI engineering team from the phone.

So the general theme of early feedback from VCs is more customer validation (fair). So if you have a 9-5 in a big org, if you’re risk team can’t answer what agents do we have and what are they doing; hook me up. It’s where I’m going to be spending a lot of my time by the looks of it.

There is a lot of chat about what the right interface is for coding agents, I figure I'd share my thoughts on this. The beauty of software being as readily available as it is means that you should be able to design the interface that works for you. For me this means: - No CLI (I just don't enjoy looking at it for long periods) - Mobile first (Running the startup means I'm on the go a lot, I'm not tied to my machine for supervision) - Focus on specs and plans more than implementation. This has been the big focus since Ralphing (Cheers @GeoffreyHuntley) became a "thing". Ive built back-pressure in more agents that review the implementation and adherence to Acceptance criteria, this gives me much higher confidence in the eventual output. What this means for me: - setting up the project and getting strong guidance docs in there, agents.md etc = high time/ effort investment. - Defining the scope or section of work we will do = high time/effort investment - Reviewing the plan, refining the plan = moderate to high time/effort investment - Monitoring execution = Low time/effort investment. Lots of time spent iterating on the review agents etc results in me being able to operate in a reactive sense during execution via push notifications. Horizontally scalable across the cluster (bumped the memory for some of the r730s today), and can run numerous scopes over numerous projects simultaneously. UI looks a bit like this:

I’ve received a lot of messages about tokenomics, buybacks, burns, staking, etc. I’ve also spoken with people who have far more experience than me in those mechanisms. This post is simply to say: I’ve heard you, and I’m not ignoring the requests. That said, my position hasn’t changed: - I will not personally manage this coin (buybacks, burns, market ops, “supporting price”, etc.). - I’ve said this from day one, and I’m sticking to it for safety transparency reasons. - The only way I intend to impact the token is by shipping real work: GRASP / ryora.ai progress and, if it makes sense, future utility. I understand why some devs choose to directly steward a token (and I respect it). Different projects, different operating models. My work is enterprise-focused, and that requires a different approach and timeline: months/years, not days/weeks. I get that some of you are trading short-term and the current price action is painful. I’m not dismissing that. But I’m also not going to cross the line into running token mechanics. My commitment remains simple: - keep building - keep shipping updates - keep learning/investigating where web3 could genuinely fit here I have some ideas, but they’re future-anchored enough that it would be irresponsible to pitch them as “soon” or use them to influence price today. If you’re here for the build and long term value: you’ll keep getting progress. If you’re here for active token management: that’s not something I’m going to do.

I think @steipete’s take on ai bots in human centric channels is interesting. Mate if you see this I wouldn’t mind your take. Is your hesitance or opposing position based on: - lack of confidence in bot quality resulting in high noise to signal ratios? Or - these environments have a ethical aspect where we should be expecting and valuing human to human interaction even if that may be more inefficient in answering questions?

I enjoyed this far too much. I find my loop is: Expand Optimise Expand Optimise

This is where we are up to, the browsers are significantly higher effort to get them to work with $GRASP but testing is promising. Ive kept chipping away at some of the more popular coding CLI tools and I've started with the IDEs too. Gotta $GRASP em' all

Is anyone in my network super into AI browsers? I fell like my control expectations aren't crazy but it looks like browsers are a generation or two behind the CLI tools. I could be missing some of the options by not having an enterprise account but from my initial testing; @OpenAI Atlas is effectively a free for all (no real config options) and no ability to add extensions. @perplexity_ai Comet looks a bit better, block certain domains etc , and you can install extensions. @Sigma_Browser looks like Atlas in terms of behaviour config, but it does allow extensions. @diabrowser was neither configurable or extensible. @genspark_ai I couldn't find any config settings for the AI but it did allow extensions. @brave had some nice general AI config options but nothing in the control realm, does allow extensions. @MicrosoftEdge Copilot was also just basic AI preferences, very little around control. Does allow extensions though --- Still to test: @opera neon <- mandates I pay before I can even download the platform bleaurghh @flowith wouldn't allow me to use my company email address? then once you finish the signup you also need an invite key? This was an annoying waste of 5 minutes. It also played funky with the Mac UI. --- Am I going insane though, when I think of the potential GRASP analysis of browser agents: G - Very little governance, install and off you go, there is SOME observability but hard to differentiate user interactions and AI interactions. R - THE INTERNET, ALL OF YOUR STORED CREDENTIALS A - Varied, depends on the user. S - extremely varied, because the reach is so broad its hard to pin down what the failure scenarios are and thus what safeguards are appropriate P - Again, huge spectrum The TLDR is unless you can enumerate the bounds by which an AI Browser will operate the unknowns are so broad it's genuinely debilitating. I'd love to connect with some browser experts though and understand what im missing here. --- Shout out to @LayerxSecurity who actually had a decent write up of a bunch of the browsers. If you guys get this notification I wouldn't mind having a deeper chat.

Its only fun when the agents play together

Clearest indicator that someone ‘doesn’t get it’, how you could hire for this role in 2026 baffles me