Your real-world agentic AI setup
Discussion
I could not see another thread on Agentic AI so wanted to start one to see what others are using and innovative ways they are using it. I am still very much at the information gathering stage, exploring limits and just messing around with this on personal projects but it is certainly the most fun I have had in awhile. Genuinely, a few times an hour I utter a 'oh f***' in amazement.
I have setup a new development team and due to hardware constraints I have limited it to 5 members who have names and roles but I want to keep that bit secret as they are my team and I created them
The lead engineer orchestrates and delegates tasks from the task list to the appropriate engineer who has their own responsibilities, strict instructions and guard rails. Once the task is completed and all tests run then they message back that they are ready for other work. I don't currently have a dashboard overview for all this as it is just in terminal but I might get the team to work on that at some point.
Currently using Claude Code Max but this stuff is moving so fast I am already considering moving over to Codex or maybe an OpenCode setup so I can use cheaper models according to task.
One thing is for certain is that I need a much more powerful MacBook going forward and with memory prices escalating this is going to hit the credit card pretty hard. I might explore some options to delegate some tasks to a VPS.
What is your setup?
I have setup a new development team and due to hardware constraints I have limited it to 5 members who have names and roles but I want to keep that bit secret as they are my team and I created them
The lead engineer orchestrates and delegates tasks from the task list to the appropriate engineer who has their own responsibilities, strict instructions and guard rails. Once the task is completed and all tests run then they message back that they are ready for other work. I don't currently have a dashboard overview for all this as it is just in terminal but I might get the team to work on that at some point.
Currently using Claude Code Max but this stuff is moving so fast I am already considering moving over to Codex or maybe an OpenCode setup so I can use cheaper models according to task.
One thing is for certain is that I need a much more powerful MacBook going forward and with memory prices escalating this is going to hit the credit card pretty hard. I might explore some options to delegate some tasks to a VPS.
What is your setup?
Sorry, this is irrelevant to your question, but as someone who has absolutely no requirement for any AI assistance, can I ask what you’re doing with the AI?
I keep being told I need to embrace it but I have literally zero need for it. I haven’t owned a computer for fifteen years now so I’m struggling to think of a single way it would improve my life - hence why your thread piqued my interest as I’m trying to find what I’m missing.
I keep being told I need to embrace it but I have literally zero need for it. I haven’t owned a computer for fifteen years now so I’m struggling to think of a single way it would improve my life - hence why your thread piqued my interest as I’m trying to find what I’m missing.
At this stage just learning the possibilities.
Many workplaces have not embraced the agentic stuff yet so like many other developers I am trying to get ahead of the game and seeing what it can do, what it cannot and assessing best practices, pain points and risk. Putting your head in the sand is not an option at this stage.
But like the saying "Everyone has a book in them" this is also probably true of apps. A couple of years ago it would cost a fortune to hire a team to develop an application. It is now quite possible to do so with your own little team of agents who are tireless and can work 24/7. It is also best to learn this stuff whilst the big labs are offering cheap tokens. The prices will ramp up soon enough.
Many workplaces have not embraced the agentic stuff yet so like many other developers I am trying to get ahead of the game and seeing what it can do, what it cannot and assessing best practices, pain points and risk. Putting your head in the sand is not an option at this stage.
But like the saying "Everyone has a book in them" this is also probably true of apps. A couple of years ago it would cost a fortune to hire a team to develop an application. It is now quite possible to do so with your own little team of agents who are tireless and can work 24/7. It is also best to learn this stuff whilst the big labs are offering cheap tokens. The prices will ramp up soon enough.
Crumpet said:
Sorry, this is irrelevant to your question, but as someone who has absolutely no requirement for any AI assistance, can I ask what you re doing with the AI?
I keep being told I need to embrace it but I have literally zero need for it. I haven t owned a computer for fifteen years now so I m struggling to think of a single way it would improve my life - hence why your thread piqued my interest as I m trying to find what I m missing.
Assuming (from the "haven't owned a computer for 15 years") part you are not running a business? I keep being told I need to embrace it but I have literally zero need for it. I haven t owned a computer for fifteen years now so I m struggling to think of a single way it would improve my life - hence why your thread piqued my interest as I m trying to find what I m missing.
For me, it's looking at agentic AI as a way of streamlining processes in our business (discussed elsewhere on the "How do you use AI" thread). One of the things that came out of that thread is that many people don't realise how much it can do so can't see where they would use it. A lot of people see it like an extra-helpful Google, which is to massively miss the point of its capabilities. Basically, anything a human can do (well OK, almost anything...) can be done by an AI agent, usually quicker and cheaper.
Beware of scammers using AI to scam people
"Conjoined twin 'influencers' revealed to be AI"
https://www.dailymail.co.uk/galleries/article-1552...
"Conjoined twin 'influencers' revealed to be AI"
https://www.dailymail.co.uk/galleries/article-1552...
Is the quality of outcome better than delegating to the other version of AI - Actually Indians? Feels like by the time you've defined it adequately and checked it and fixed it you're no better off?
Straight DevOps automation is a different thing, mostly becauses its simple, repetitive and easily defined.
If I want to hand off dev tasks these days I like using Romanians, they have actual intelligence and can take a rough outline & run with it without too much supervision or checking, and don't have an innate desire to tell you what will keep you happy or pretend to understand.
Straight DevOps automation is a different thing, mostly becauses its simple, repetitive and easily defined.
If I want to hand off dev tasks these days I like using Romanians, they have actual intelligence and can take a rough outline & run with it without too much supervision or checking, and don't have an innate desire to tell you what will keep you happy or pretend to understand.
JoshSm said:
Is the quality of outcome better than delegating to the other version of AI - Actually Indians? Feels like by the time you've defined it adequately and checked it and fixed it you're no better off?
Straight DevOps automation is a different thing, mostly becauses its simple, repetitive and easily defined.
If I want to hand off dev tasks these days I like using Romanians, they have actual intelligence and can take a rough outline & run with it without too much supervision or checking, and don't have an innate desire to tell you what will keep you happy or pretend to understand.
If PH had a 'like' button I would have pressed it for this post.Straight DevOps automation is a different thing, mostly becauses its simple, repetitive and easily defined.
If I want to hand off dev tasks these days I like using Romanians, they have actual intelligence and can take a rough outline & run with it without too much supervision or checking, and don't have an innate desire to tell you what will keep you happy or pretend to understand.
It's quite naughty.
AI is now 'Actually Indians' in my head forever.
You are a bad man.
harryt said:
If PH had a 'like' button I would have pressed it for this post.
It's quite naughty.
AI is now 'Actually Indians' in my head forever.
You are a bad man.
Not mine originally, there's a long inglorious history of people like Amazon and others claiming to be using 'AI' for tasks that wasn't exactly what it was assumed to be.It's quite naughty.
AI is now 'Actually Indians' in my head forever.
You are a bad man.
And I've been around the loop myself of spending lots of time trying to hand off work to every possible thing that becomes briefly popular.
272BHP said:
At this stage just learning the possibilities.
Understood. I’d like to not fall behind too far so I suppose that’s why I’m trying to take an interest in it. Maybe I’ll even spot an opportunity. FlyingPanda said:
Assuming (from the "haven't owned a computer for 15 years") part you are not running a business?
For me, it's looking at agentic AI as a way of streamlining processes in our business (discussed elsewhere on the "How do you use AI" thread). One of the things that came out of that thread is that many people don't realise how much it can do so can't see where they would use it. A lot of people see it like an extra-helpful Google, which is to massively miss the point of its capabilities. Basically, anything a human can do (well OK, almost anything...) can be done by an AI agent, usually quicker and cheaper.
No, I don’t run a business, I drive aeroplanes. I suppose I can see far more sense in it if running a company. In a way in my job we’ve been handing over control to the machines for decades but from a professional point of view I’m struggling to see how I can make use of it. I tried to get ChatGPT to summarise a flight briefing but uploading it and then reading the summary saved me about a minute. It’s in my own time and my own life that I’m trying to see how it could improve things and to try and see the possibilities. For me, it's looking at agentic AI as a way of streamlining processes in our business (discussed elsewhere on the "How do you use AI" thread). One of the things that came out of that thread is that many people don't realise how much it can do so can't see where they would use it. A lot of people see it like an extra-helpful Google, which is to massively miss the point of its capabilities. Basically, anything a human can do (well OK, almost anything...) can be done by an AI agent, usually quicker and cheaper.
I’ll follow with interest….
JoshSm said:
Is the quality of outcome better than delegating to the other version of AI - Actually Indians? Feels like by the time you've defined it adequately and checked it and fixed it you're no better off?
Straight DevOps automation is a different thing, mostly becauses its simple, repetitive and easily defined.
If I want to hand off dev tasks these days I like using Romanians, they have actual intelligence and can take a rough outline & run with it without too much supervision or checking, and don't have an innate desire to tell you what will keep you happy or pretend to understand.
I mimicked a feature locally that took a junior developer 4 weeks to complete successfully. The agents did it in about 2 mins, including review, documentation and comprehensive tests. It also decided early on that the instructions in my prompt was perhaps not the best way to go and suggested another way of doing it that was innovative, more testable and just better in every way.Straight DevOps automation is a different thing, mostly becauses its simple, repetitive and easily defined.
If I want to hand off dev tasks these days I like using Romanians, they have actual intelligence and can take a rough outline & run with it without too much supervision or checking, and don't have an innate desire to tell you what will keep you happy or pretend to understand.
This is Feb 2026 what will they be capable of in a years time?
Instead of "more powerful Mac book" you might want to consider building a PC-based AI server and offloading the processing.
Bit of a problem right now because of RAM prices, but I've seen people skip the Google Coral TPU and go straight to multiple GPUs and oodles of RAM on their home "neural server" machines. I haven't done this as I currently don't have the need, but can see myself going this route when things are more sensible again.
Alternatively you CAN rent processing power but when I investigated this recently it was a pretty expensive way to do things if you had an ongoing need.
Bit of a problem right now because of RAM prices, but I've seen people skip the Google Coral TPU and go straight to multiple GPUs and oodles of RAM on their home "neural server" machines. I haven't done this as I currently don't have the need, but can see myself going this route when things are more sensible again.
Alternatively you CAN rent processing power but when I investigated this recently it was a pretty expensive way to do things if you had an ongoing need.
We’re a bit more locked down where we are but I can use agentic Claude through VS Code
Today I wrote a script that updated Azure tokens and triggered some actions on Databricks and another vendor app, calling about 10 different REST/graphql apis
It took me less than 2 hours, I think a year ago that would have been a day or two’s work
The biggest timesaver was its ability to pull the info from API responses, that’s always a fiddle in Python. Also it was also able to parse and reuse proprietary code (modules etc) already in use in the codebase, just off simple prompting, which was really impressive.
Exciting stuff.
Today I wrote a script that updated Azure tokens and triggered some actions on Databricks and another vendor app, calling about 10 different REST/graphql apis
It took me less than 2 hours, I think a year ago that would have been a day or two’s work
The biggest timesaver was its ability to pull the info from API responses, that’s always a fiddle in Python. Also it was also able to parse and reuse proprietary code (modules etc) already in use in the codebase, just off simple prompting, which was really impressive.
Exciting stuff.
donkmeister said:
Instead of "more powerful Mac book" you might want to consider building a PC-based AI server and offloading the processing.
Bit of a problem right now because of RAM prices, but I've seen people skip the Google Coral TPU and go straight to multiple GPUs and oodles of RAM on their home "neural server" machines. I haven't done this as I currently don't have the need, but can see myself going this route when things are more sensible again.
Alternatively you CAN rent processing power but when I investigated this recently it was a pretty expensive way to do things if you had an ongoing need.
I need a new Macbook anyway so that has to be the first purchase. I will explore a home server though, quite the rabbit hole I am sure Bit of a problem right now because of RAM prices, but I've seen people skip the Google Coral TPU and go straight to multiple GPUs and oodles of RAM on their home "neural server" machines. I haven't done this as I currently don't have the need, but can see myself going this route when things are more sensible again.
Alternatively you CAN rent processing power but when I investigated this recently it was a pretty expensive way to do things if you had an ongoing need.
I love using AI at work to simplify laborious tasks.
Instead of receiving a dozen hefty design documents , along with a 30 page contract with various delivery milestones and making a project plan, RACI, timeliness, BOM etc,
Load it all up and ask AI to do it all and then spend an hour proof reading it. It does make mistakes, but doing it the old fashioned way I always made mistakes as well. But instead of spending a day writing it, then an hour proofing it, I cut out the boring bit :-)
I have also started getting it to write contract change notes and project change requests when it picks up issues between the project delivery in practical terms and what's in the contract.
Instead of receiving a dozen hefty design documents , along with a 30 page contract with various delivery milestones and making a project plan, RACI, timeliness, BOM etc,
Load it all up and ask AI to do it all and then spend an hour proof reading it. It does make mistakes, but doing it the old fashioned way I always made mistakes as well. But instead of spending a day writing it, then an hour proofing it, I cut out the boring bit :-)
I have also started getting it to write contract change notes and project change requests when it picks up issues between the project delivery in practical terms and what's in the contract.
I've used Codex briefly and it seems pretty good.
Most of my stuff is just done in normal ChatGPT. I give it the outline of what I want, it asks a few questions, and posts up the code I need. I go back with any problems, changes or bugs and it regenerates what I need. I've written a helpdesk, AI driven email processor, lots of little utils (spots new IPs on the network, filters jobs). Less than an hour and I've working code. All this is greenfield stuff, but using Codex above I could work on existing code base efficiently.
Nothing Agentic as such yet though.
Most of my stuff is just done in normal ChatGPT. I give it the outline of what I want, it asks a few questions, and posts up the code I need. I go back with any problems, changes or bugs and it regenerates what I need. I've written a helpdesk, AI driven email processor, lots of little utils (spots new IPs on the network, filters jobs). Less than an hour and I've working code. All this is greenfield stuff, but using Codex above I could work on existing code base efficiently.
Nothing Agentic as such yet though.
The trouble is that this stuff moves so fast.
My little team of agents have now been fired as they have been superseded by the new agent flow in Claude that was released yesterday. I can now spawn agent teams according to task.
Watching them argue with each other and then eventually come to agreement on how to proceed is fascinating but a little bit unnerving as well
How powerful is this stuff currently?
https://www.anthropic.com/engineering/building-c-c...
My little team of agents have now been fired as they have been superseded by the new agent flow in Claude that was released yesterday. I can now spawn agent teams according to task.
Watching them argue with each other and then eventually come to agreement on how to proceed is fascinating but a little bit unnerving as well
How powerful is this stuff currently?
https://www.anthropic.com/engineering/building-c-c...
Anthropic said:
To stress test it, I tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kernel. Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.
I am using Claude for my own personal projects and it's pretty good - but validate everything at every step.
A lot is down to the prompting and ability to test - I think it's a perfect sidekick to dev work (nope, never Copilot).
Yesterday for example I wrote a mongoDB stack for OpenCanary in 40 minutes from scratch. I invested another 30 minutes to make it more "humanlike" insofar that it mimics a deployment without authentication facing the Internet and allows data "destruction" and ransom note-drop,
Would I replace a team of qualitied devs with it? Nope - but the deve I deal with would have loved it for example when they had to review an SDK for our mobile banking app some years ago - 10000 lines of crap code assessed in 1 minute? Yes please.
A lot is down to the prompting and ability to test - I think it's a perfect sidekick to dev work (nope, never Copilot).
Yesterday for example I wrote a mongoDB stack for OpenCanary in 40 minutes from scratch. I invested another 30 minutes to make it more "humanlike" insofar that it mimics a deployment without authentication facing the Internet and allows data "destruction" and ransom note-drop,
Would I replace a team of qualitied devs with it? Nope - but the deve I deal with would have loved it for example when they had to review an SDK for our mobile banking app some years ago - 10000 lines of crap code assessed in 1 minute? Yes please.
Gassing Station | Computers, Gadgets & Stuff | Top of Page | What's New | My Stuff