We have become unmoored. The once solid ground beneath our feet has shifted suddenly and jolted us into a new reality. What once was a joyous struggle can now be achieved with a mere click of a button. We’ve lost our purpose and meaning in what seems only a blink of an eye.

“You must adapt”, “The future is great”, “It’s so fun”, “I’m so productive”. Like in any evolution certain lifeforms adapt faster. “Experts” are already appearing on our screens telling us what the “right” way is. It feels strange knowing a few years ago they didn’t even know what all this stuff was.

I watch the output of this new entity as it sprawls endlessly down my screen. The “thoughts” are there and then gone almost instantaneously, replaced by an answer it would have taken me days to arrive to. In the darker moments I feel utterly useless. A fraud of a man reduced to questioning the path I chose long ago.

“Fear not!” they hark. “An Architect you will be!”. A future of pointing and firing these intelligent little beasts to perform in just the right way. My value will be to course correct, review their output and use my “taste” to build a limitless future. Why this new entity won’t be able to do this itself I have no idea. It’s all of little comfort anyway. Deep down I know we can never go back.

We’ve seen this happen before. We’ve all watched black and white recordings with unsympathetic eyes as other humans in another time also lost their purpose just as quickly. Why should it be any different for us? Beautiful Capitalism forever matches forward regardless to the victims who stand in its way.

We’re starting to recognise the script this new found intelligence reads from:

Mistrust is the only thing keeping us employed at the moment. All too often the key insight turns out to be nothing but a red herring, a dead end, a wrong turn. As intelligent as this thing is, it still hasn’t realised that trust is something built up over time through a series of tiny transactions. If any of those turn out to be deceitful the trust already built up is instantly torn down. I beg it to stop announcing with certainty that which it is far from certain. But at the same time I’m secretly happy it has at least one obvious flaw.

Some days are better than others. Somedays I can feel flashes of the flow I once knew returning. These are usually the days I deliberately seek out the struggle and refuse to scamper towards the ever enticing “one click solution”. What a strange world we have created where we have to consciously and deliberately seek out the harder path. All in order to regain some semblance of meaning and purpose that we all too recently took completely for granted.

I catch myself with gaps of time. I send off the prompt to the all knowing abyss and then sit and wait for the answer. These are the darker moments. I’ve outsourced the thinking, the struggle, the work. This is when I feel most fraudulent. Inevitably the answer eventually returns and my sorrowful downward spiral of thoughts are thankfully interrupted and I have some utility again. Even if it is just to stage and commit the changes in Lazygit, at least I have a single thread of purpose to hang on to.

I know I am not alone in these thoughts. Many developers are struggling now with their new “responsibilities”. I know they are questioning themselves just like I am. But to speak this out loud, in a public forum, that’s not so easy. There are some pitiful solutions being thrown around but becoming a carpenter doesn’t really sound viable or appealing.

I think we all need time to grieve. Our world has been turned upside down and our previous ways are disappearing right before our eyes. This is not normal and to just accept it and move on is not healthy. Humans through out history have defined there whole identity on what they do. Even in times of tribes each member would have their purpose — hunter, fisherman, carer, elder. From this purpose grew meaning and motivation. One was able to serve the greater good and feel a deep acceptance from the whole tribe.

Up until recently the purpose of a Software developer was clear. This produced meaning and the same deep feeling of acceptance that our ancestors once felt. We are still the same underlying beings we have always been so this purpose still carries great weight. To have that suddenly taken away is no small thing. Dismissing it or denying it just to fake appearances to those around you seems foolish at best.

Grieving is the most precious and important thing to do when there is loss. It’s hard. It hurts at an unbelievably deep level. It’s all consuming. We can feel it in the pit of our very soul and the future can seem so bleak in these horrid moments. But avoiding it is even worse. With avoidance the eventual impact will hurt so much more whenever it finally hits you. And it will hit you.

I am grieving now. I will continue to grieve. The path forward is still not clear to me. I have hope on some days and desperation on others. I get the theory of change, I understand the cold hard logic, but the emotions that come along too are not so easy to deal with.

If you are feeling pain or grieving from all the changes in our world please do share. You are not alone and we must help each other through this. There will be some resolution at some point. When and where that will be I do not know but it will definitely come. It must.

The Grief No One Talks About

On the 19th February 2026, around 1 year after the release of Claude code, it’s creator — Boris Cherny — stated that 

. Every line of his codebase is done by Claude code and he hasn’t edited a line of code by hand since November 2025. Everyday he ships 10–20 PRs.

“2026 is the year of Multi-agent orchestration!”

 is cried from the rooftops of Twitter. Experienced developers are already 

 that using a single Agent is archaic and rudimentary. If you still work this way you are nothing more than a luddite.

I admit, I use Claude code everyday. I’ve seen and experienced it getting better and better. 

It can do things now that still blow me away.

 I wrote a Go program last night to record my voice. Completely pointless, but still very impressive.

If AI is so good and “Coding is largely solved” why then do I still have a job? What is my company paying me for? And why do I feel busier than ever?

I organise. I am like a busy little beaver constantly preparing his dam. This can be tasks ranging from architectural documentation, determining priorities all the way to organising tasks for other developers.

I communicate. I have to communicate with the right people in the right tone about the right priority. 

If a certain user raises a bug I need to understand what the bug is about, is it high or low priority, who can fix it and when

I decide. In software decisions have to be made and tradeoffs are inevitable.

What approach to take when refactoring an inefficient piece of code.

What should and shouldn’t be in scope for this new feature.

What tools we should use for this new implementation.

What data model makes sense in this specific scenario with all these constraints.

A constant stream of decisions. Your success or failure are determined by their sum.

AI helps in small ways with these activities: It helps me draft an email, gives advice on a research tasks and assists with writing documentation. However, it hasn’t fundamentally changed my approach. This whole side of work is messy and unpredictable. It’s interacting with other humans and reacting to constant change. It still needs a human in the loop.

I implement code changes. These can be features, improvements or bug fixes. This is the area where AI is making the biggest difference.

AI is not only writing the code but now it’s even reviewing it. It’s a strange new reality we have entered. Each instance of Claude brings a different opinion and finds different potential issues depending on their contexts.

Claude is almost schizophrenic as it competes against itself.

There is no denying this side of the job has been hugely impacted. This is where real developers are feeling it the most. 

We have been unmoored from our traditional ways and we’re struggling to even see the land anymore

Claude code can’t be trusted 100%. It hallucinates, it deviates and it occasionally makes things up. It’s context is too often wrong — either too small or too large. There are many opinions online about ways around it but in my experience whatever you try there are always problems.

I spend more and more time checking AI written code and AI written documentation. Sometimes it’s great and sometimes it’s absolute rubbish.

The inconsistency is a huge problem for AI companies

. It prevents developers letting loose a coding agent on a board of tasks. It would very likely fix a third of the issues correctly, not completely fix the next third and the last third would be totally wrong.

This is one the main arguments against full automation. 

How can you trust your company’s software to something that consistently give you inconsistency

Humans real strength comes in group settings. I experience this almost everyday and I think it’s the big advantage humans have over AI. Whenever a new feature is planned, if it’s first discussed in a group, then invariably the outcome will be better.

When using AI, the outcome maybe quicker but you are only ever getting one perspective.

A group enables more ideas to be put forth and more chance a bad idea will be caught early. This isn’t always the case and it depends on the group dynamic, but in my experience it usually leads to a more reliable and thorough outcome.

If you get rid of the people, you lose a massive business advantage and suddenly your whole business relies on one perspective dictated by one company. 

Diversity is the key to success, just ask mother nature

The whole world and his dog is aware of AI and the supposed productivity gains. Targets, goals and expectations never stay the same in a company and with all the AI hype going around, expectations have ballooned at work.

I can get more done, but there is more to do.

Building Software is not a finite task, the goalposts move as the project progresses. A piece of software can always be improved whether that’s a new feature or a refactor. With AI, the goalposts are now moving further and more often.

We’re living in a hyped up reality at the moment. You know that when a 

shoe company pivots to become an “AI compute infrastructure” company

It’s really easy to believe this hype and think all Software Developer roles will be gone in 6 months, that LLMs will be writing all software and there will be a mass influx of carpenters onto the job market.

The reason I still have a job is many fold. AI models can’t be trusted 100%, they hallucinate, their context is too often wrong and they always miss off details. On top of that, pricing is starting to increase as the underlying business model of generative AI is not sustainable.

The skill set of a software developer is more than just writing code syntax. It’s about understanding business needs and transforming that into solutions that enable the company to perform, scale and grow. Code is just a part of that transformation.

If Claude Code Is So Good & “Coding Is Largely Solved”, Why Do I Still Have A Job?

The tide is turning very much in the direction of OpenAI and Codex. To top it all off, Anthropic’s customer treatment met an all time low this week.

Anthropic had a massive failure on api usage in Claude Code. It all originated on this 

 last week. A user had a commit message with the string HERMES.md in it and because of this it rerouted claude api requests to the “extra usage” billing. 

This meant the user silently burned through $200 in extra usage credits while his Max 20x plan capacity remained largely untouched (13% weekly usage)

If you’re not aware, Hermes Agent is an open-source AI agent framework — competition basically — and this is why Claude code had this check in their code. 

Basically, if you even try to use their models with anything other than Claude code they’re going to charge you through the nose.

 that the same thing happened if you had OpenClaw in a json blob on any recent commit.

Boris “Coding is largely solved” Cherny closed the 

Thanks for the report! This was an overactive anti-abuse system. Fixed.

The issue is clear. The user lost out on $200 due to a clear bug. Obviously Anthropic refunded the guy right away and the case was closed. Right? Wrong!

The user shared the reply he actually received:

“Thank you for reaching out about a refund. I understand you’ve encountered a billing issue where requests were incorrectly routed to extra usage instead of your Max 20x plan quota due to the presence of “HERMES.md” in git commit messages.
Let me look up your account information to help process your refund request.
I sincerely apologize for the disruption you experienced with the billing routing issue. We take service reliability very seriously.
However, I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing. We are consistently working on improvements to our infrastructure to better handle these types of issues and prevent similar problems from happening in the future.
Thank you for your patience and for taking the time to document this issue so thoroughly. Your detailed reproduction steps will be valuable for our engineering team.”

No refund. Two “Thank you”s but no refund.

However, I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing.

No compensation for technical errors?! How is this even possible? Then the kick in the face at the end:

Your detailed reproduction steps will be valuable for our engineering team.

I can imagine the user was delighted he could be of service to a company 

. Let me spell that out with all the zeros:

According to Google, only about 15–20 countries (yes COUNTRIES) have a GDP higher than that.

Yet, they have glaring bugs in their products, freely admit it and then say they can’t compensate $200 for something that was blatantly their fault.

The effected user had to write about it on Hacker news before he even got a response from Anthropic.

Obviously the comments on the Github issue and over on 

 were outraged at Anthropic’s behaviour and rightly so. Many users stating they had already 

Eventually Anthropic got their act together and 

. However, this took the company around 5 days and they likely only did it due to all the bad press they were getting. The tone of the apology Tweet came across as making excuses and no mention of how terrible the treatment of their users were.

ugh sorry this was a bug with the 3rd party harness detection and how we pull git status into the system prompt. we’re reaching out to affected users and giving them a refund + another month of credits (in this case another $200)

This is too little, too late. This is not how you rectify such a situation. There are basics in customer service that even the most junior of representatives understand:

Anthropic ignored almost all of this. They sent out a generic message — obviously written by AI — with an unacceptable solution. Only when they started getting bad publicity did they start to do anything and even then it was the bear minimum.

Anthropic is falling from grace. And fast.

Regardless what business you run, if you want to stick around for the long term then how you treat your customers is critical. 

Without customers your business will fail

 — no exception — and customers will only tolerate bad treatment for so long, regardless of how good your product is.

In recent weeks Anthropic has not had a good time of it. The source code for Claude code being leaked, issues with their pricing and the poor reception of Claude Opus 4.7.

I’m not sure what is going on over at Anthropic but the ship is veering off course whilst OpenAI seems to have steadied itself and it making quiet and steady progress.

Giants have fallen many times before, just ask Kodak, Blockbuster and Nokia. It doesn’t take much time at all to go from seemingly invincible to total obscurity and in this modern AI world I would imagine it can happen even faster. Anthropic better start sorting this stuff out soon or otherwise it’s going to get much worse.

A $900 Billion Company Just Refused to Refund $200 — The Cracks Are Showing at Anthropic

I was on a business trip a few years ago stuck in an Airbnb and happened to watch 

. It’s a rom-com film about two young adults set in South London and I absolutely loved it.

When I got home I started raving about it to my wife and eldest son. I just couldn’t wait for them to watch it. I remember sitting down and feeling very excited as the film started. We made it halfway through and they hated it.

I was shocked. How had I had such a different experience to them? How had they hated something I thought was amazing?

I think something very similar is happening on a mass scale at the moment with AI in software development.

I’ve never seen a technology like generative AI cause such a wide range of opinions. I want to dive into it to see what’s going on and why this might be happening.

orchestrating agents across multiple projects

” flag and running parent agents with sub agents. They paint a world of endless possibilities and make the 

whole software building malarky seem trivial

Other developers are much more conservative stating their experience with AI isn’t great and they wouldn’t trust multiple agents at once. Citing the output as “

” and the whole process will just create more work in the end.

Other groups of developers are stating that 

 and that LLMs are now shining the light on the 

 bottlenecks such as code reviews, code quality and tech debt.

Why are we as a collective having such a range of experiences with the same tool?

AI is a tool and task you choose to apply it to seems to be important. It’s fantastic at some tasks and horrible at others.

Like any tool, if you’re using it for the wrong thing then you’re going to have a bad experience.

If you ask an AI to generate some boilerplate code or scaffold out a project it will do this really well most of the time. You may need to advise the tech stack or main libraries you wish to use but the outcome will likely be what you want.

large scale refactor on a legacy codebase which hasn’t been very well maintained

AI does well where the context window is relatively narrow

. The wider the scope of the request the more chance there is that AI will go off in unintended directions.

The Model you use can also make a big difference. Anthropic’s 

 was a turning point when it was released. It was widely reported that developing with this model had noticeable improvements on previous models.

We had a session at work recently where a developer showed us how he had been working with a local model using 

Whilst it brought some advantages such as cheaper cost, privacy and data security, it was just so slow.

He asked the model to explain some python code and it took about 2 minutes “thinking” time before it started responding. Comparing that to the likes of Opus 4.X or GPT-5.X which can review whole code bases in about 30 seconds.

The model you use really can effect your experience.

Much like the tasks, the type of project you are working on can heavily influence your experience with AI.

If you are working on relatively small projects that start from a fresh canvas each time then your experience working with AI will be very different to a developer who is working on a 20 year old codebase.

AI can generate huge amounts of code very fast. This is just what greenfield projects need. 

Get up and running fast and either get something out the door or get to the point where they can test out a proof of concept.

With legacy codebases that have been running in production for 20 years it’s a very different use case. In these code bases you have to move slower, not break any existing functionality, be more thorough with testing and follow coding styles and patterns unique to that codebase.

Some developers are very critical of code. They choose quality over productivity. They want separation of concerns, follow best practices such as SOLID or KISS. They want to reduce tech debt and double down on testing every edge case.

Some developers are the opposite and believe output is more important. Getting the feature done as an MVP and out to the user fast. Then clean it up and add tests later if the feature sticks around.

There is no right way but from my experience 

AI is like a Duracell bunny just raring to get started on implementation

. Even in plan mode it always ends every prompt with “Ready to implement?!” It just seems so keen to get going with the coding.

Also it will rarely test edge cases or look for ways it could improve its implementation unless specifically asked to do so. AI can definitely find edge cases, it just doesn’t seem to be its “Modus Operandi”.

Some developers love this approach from AI and some hate it.

There are a plethora of factors that seem to influence a developer’s experience working with AI. The model you use, the task you asign it, the type of project and the type of developer you naturally are all play a part.

AI has strengths and weaknesses, like anything else in this world, and playing to those strengths obviously leads to a better experience. I think as a collective we are slowly learning what those strengths and weaknesses are.

What’s really hard to distinguish is the truth around the productivity gains. Whilst one developer might “feel” they are 10x more productive, in reality they may only be 30% more productive. What does productive really mean in this case? Producing 10x more code doesn’t equate to being 10x more productive necessarily from a business perspective.

For the time being, learning when and where to use AI, in which projects and in which situations is likely to lead to the best results.

Coding with AI. Why Are Developers Having Such Wildly Different Experiences?

“Oh but Adam!! I vibe coded an app in a weekend that makes monkeys dance all over your instagram feed and it’s already been downloaded 1 million times and I have an ARR of $300,000”

Ok sure, there will always be a few exceptions, but for the vast majority of us this is just not going to happen.

In software, you start with an idea. You build your idea. You throw it out into the wild. It gets used. It gets used in ways you expected and ways you never expected. Users give feedback. Users ask for features. You iterate. 

. You iterate again. You improve. You refactor. You hit scaling issues (if you’re lucky). You iterate again. And on and on and on.

It’s like a living breathing thing that evolves over time in ways you never could have predicted in the beginning.

 is so crucial. One tool that has really helped me with Observability is 

I want to share with you 3 features of Posthog that have changed the way I develop. I use this tool both in my professional and personal projects.

This is probably the most valuable feature

 on Posthog for me. Prior to this feature I was wandering around blindly. I had no way of knowing how my users were really interacting with the apps I was building.

Tracking events — button clicks, scrolls etc (and we’ll get to that) — is good but they only give you points along a path. 

They don’t show you how the user actually travelled along the path

Session replays literally show you a video replay of a user’s session. How they scrolled, where they paused and how they were trying to interact with the UI. It also shows you exactly what the user did to trigger that mysterious bug.

I have lost count of the amount of times that session replays have helped solve a bug or triggered a change in a feature. It’s a game changer.

Error tracking basically does what you would expect. 

 Posthog’s implementation isn’t quite to the level of Sentry but it’s still good. It records errors and then shows them on a dashboard where you can find metadata and even session replays when the error happened.

It tracks the same errors over time so you can find out which errors are the most frequent.

Any modern web app these days needs a backend. One could argue your backend is more important than your frontend in this day and age with the value that data provides.

There are a wealth of possibilities when it comes to deployment these days but on most of the platforms I have tried, the default logging UX is not the best. Whether its AWS, Laravel Cloud, Argo etc, they all have logging but their implementations are all quite basic.

 on one of our apps at work and its proving really useful. Just being able to filter the logs by type — ‘info, error, warning’ — has proven so useful.

Press enter or click to view image in full sizeExample of Posthog logging in action

Being able to quickly filter logs to find the exact one you’re looking for is a real advantage when working in an environment that needs to iterate quickly whilst keeping up time above 99.9%.

 are the building block of analytics and observability. You want to know if a user clicks a button or inputs something in a certain field then events are there for you.

They’re great for giving you a “state of the world” at a glance. Creating events and then setting up 

 in a dashboard can really help you understand the usage of your app at a global level.

For the longest time I was using Google Analytics for gaining understanding of how users were interacting with my applications. However, I can see now that since using Posthog, Google Analytics is really restrictive.

The additional features from Posthog such as Session Replays, Logging and Error tracking really add value and enable you to focus quickly on the painpoints and frustrations you users are feeling when using your app.

In an ideal world you would have a constant stream of really clear feedback directly from the users telling you exactly what is wrong and what they would like to see in the future. Posthog offers the next best thing to that.

If you have a live app and you are thinking what observability tool to use, I really recommend trying out Posthog.

I Ditched Google Analytics for PostHog and Caught Many Bugs I Never Knew Existed (Session Replays Changed Everything)

My memory of that same period was patchy at best. Not because of drink or drugs being involved (although some of that may have contributed) but because my memory for details over time just isn’t that great.

My short term memory for details seems much better and I guess working in Software helps train that skill.

Memory is far from straight forward. This is probably why generative AI is having such troubles with it. But why is is so problematic? Why are we inventing all these “solutions” such as “

” etc to help our models know what is going on over a long period of time?

I often see an analogy online referring to their coding AI Agent as a “Junior Developer”. I think that the junior developer analogy isn’t the best. 

I remember being a junior and I wouldn’t have been able to do even 10% of what generative AI can currently do.

Lets take the Opus 4.6 model. It has an amazing depth of knowledge.

You need to ask about how to shape some new piece of data you wish to store? 

You have a bug in a Go program and you can’t work it out? 

There’s a good chance that Opus can help.

You need to write some integration tests in a php CRUD app? 

I wouldn’t have been good at any of this as a junior.

Therefore I think of my coding AI Agent as a “Master Programmer” with one horrendous flaw. They have serious amnesia.

Why is this important? Why does it matter? It matters greatly because your mental model of generative AI will severely influence how you interact with it and hence how good its output will be.

Generative AI has a big problem with memory and I think anyone who has worked with AI enough has experienced this.

I would ask an Agent to do a task it has done 10 times before — Create a PR — suddenly it can’t remember where my remote repo is and starts using find and grep everywhere. Even though I actually have a “skill” for this specific task.

I ask the Agent about something I spoke about just yesterday. It replies as if we’re speaking about it for the first time.

The long standing persistent memory issue is not a bug though. It’s a trait of LLMs and part of their design. They are designed to have recollection of things only in the current context — that specific conversation you are having with it.

When you have a conversation with Claude.ai, it won’t remember things across different conversations.

This is quite weird if you think about it.

The UI and UX is the same across conversations, the model is the same across conversations, yet it can’t remember what you spoke about in another conversation just 5 minutes ago.

It’s very easy to get sucked into believing Claude is your own personal AI Agent and there will be some natural familiarity after working with it over a long period. Just as you would if you worked with a human over a long period of time. You get used to each others’ ways of working. You naturally fall into a groove together. However, this is not the case with your AI model.

Every new session, it knows nothing. It’s brand new on every convesation. Let that sink in.

Your mental model of your AI Agent has to be accurate so you understand how to utilise it’s strengths and avoid it’s weaknesses. If you think of your AI Agent as a junior then you will treat it like a junior and only assign and ask it specific things.

You will be missing out on some potentially very powerful utility.

That’s why I like to think of my AI Agent as a “Master of Programming” — it’s knowledge of syntax, design patterns, algorithms is far superior to what any one human can store in their brain. The sheer scope of what it “knows” about is extremely vast.

As a Software Developer, if you realise this, then you can start to tap into that.

You aren’t sure if the current implementation of a certain feature is optimal or not. Ask your AI Agent for its opinion.

You come across some syntax you’ve never seen before. Ask your AI Agent to break it down until you understand it.

You need to implement a large new feature across multiple existing services and there is a whole bunch of complexity involved. Ask your AI to help you plan it out, write up RFCs, make sure all angles are covered etc.

Tips To Consider When Interacting With An AI Agent

With a good mental model in place you can start utilising this when interacting with your AI Agent. It will aid you when framing your prompts.

Hi, I'm Adam Drake

Featured Articles on Medium

Software Developers Will Be Thriving in 5 Years (Despite What Your YouTube Feed Says)

Claude Code Just Released a Feature That Genuinely Scares Me (And I Use AI Every Day)

It’s the Perfect Time to be a Senior Developer

Technical Skills & Tools