The evolution of coding agents
September 2026“and as it turns out, the sand can think!”
In Nov '22, ChatGPT first came out. It was simple - you send in a message, you receive back a message. The model ran on OpenAI's servers. It was pretty dumb. It wasn't up to date and it couldn't do anything, including searching the web.
Overtime, the models started getting better, and people started to ask questions. Can we have it search the web? Can it draft this paper for me? Can it book my flight for me? Can it manage my life for me? There were a few big things that took place:
1) tool calls - People realized that you could have it output commands that you could then run, therefore giving the model the ability to take action. You could tell the model, "Hey, here is this button, and this button will let you edit a Google Doc. To call this function, send back 'TOOL: Edit Google Doc', and now you've given it the ability to edit your google doc.
So, we built a protocol around this, which we called MCP (model context protocol). And so all the software companies started building MCPs, from Expedia to construction software companies like Procore, so people could let AI use their software. Suddenly AI stopped becoming a dumb chatbot - it could take actions.
2) computer-use - Somewhere in 2024, Anthropic released a demo of computer-use, meaning the AI could click around the computer like a normal human. Before computer-use, AI was confined to using softwares with MCPs. If a software didn't have an MCP, it was very difficult to actually access it. (This is partially true, there were still ways, but very volatile). Computer-use to this day feels magical.
3) AI got really good at coding - Ultimately, all software is code, and the models got REALLY good at writing code. The reason for this is three fold - code is instantly verifiable (i.e you can run it and no if its right or wrong in real time), code is deterministic, and there's alot of freakin code. Code is a much more structured language than english. As humans, our words and language compress meanings of reality, so words don't mean the same thing to everyone. Code is different - there's no real room for debate. So the AI models got great at code, and then could write code to interact with other pieces of software. Ultimately it lives in a computer which runs on code, so 'learning' this language made it second nature for AI to interact with all the software around it.
4) Reinforcement Learning: At a surface level, you ask the model the same question multiple times, pull out the best answer, and feed it back in. This makes the model better at reasoning.
5) AI agents: we combined all of this and wrapped it into a loop - So you have this AI model that can write code, interact with software, and click around a computer. What if you just let it continuously run? Instead of the basic 'I say something, AI responds, I say something else', let it run for long periods of time. Let it try something, then let it look at the results of what it tried and go again. Encourage it to take its time and 'think', let it verify its done the right thing. At first, agents were a mess. 2 answers in and they became stupid, started saying all sorts of nonsense. So what people did was they built 'workflows', where you define exactly what the AI should do in a structured way. But as the models got better they had their forrest gump moment, and now its better to let the models run free and figure out what to do.
So here we are. We have agents that can run for long periods of time, write really good code, and interact with any type of software. The first & biggest use case for these agents were, well, to write code. Developers, companies, even non-technical people can now simply prompt these agents and they can spin up apps and test them. We've recently gotten to the point where even extremely technical developers don't even look at the code anymore. They just ask for things in english, and those things get done and tested.
Developers have an advantage in wielding coding agents, because they understand the underlying systems, but this advantage is degrading. The AI models are getting so good that in a couple years or so, its likely that we won't need human coders at all. You simply say what you want, and the coding agents will create it from start to finish. They will decide the technologies to use, the systems around the tech, etc, and they'd do a better job at it than even the best coders in the world.
But coding agents aren't just good at creating software - they write code to interact with software, making them vital for pretty much everything. The AI models run on computers, and for it to be able to interact with the computer that its housed in, it does so using code. Code is the action language for AI, just like movement is for humans. We move through space with our bodies, AI moves through code. A good analogy is video games - we control the game through our thumbs (and feet if you're playing Just Dance), AI agents control them through code. The better the code that the AI writes, the more things they can do.
What that means is, code is now an integral part of any AI agent - a tool to interact with the world. And as the AI models get better, not only can they move better & develop better software, they also just get better at things in general. So coding agents, agents, and AI are essentially interchangeable terms.
The coding agent variant of ChatGPT is called Codex. Codex to ChatGPT is what the Iron Man suit is to Jarvis. Jarvis can be incredibly smart, but the Iron Man suit gives it a surface to actually execute its prowess. You would call Jarvis the AI model (ChatGPT), the Iron Man suit the 'harness' (Codex), and together the model + the harness is the agent. Its important to understand the distinction, because there's a difference between the model and all the stuff around the model. The agent is all of it, together.
Here's the parts needed for a very powerful agent:
→ has access to several tools. the more the better.
→ has access to a great browser. This is important - it should have full control over a browser, and not be susceptible to failing.
→ can wake up / be woken up by external events (crons, triggers, webhooks)
→ identity - it should have its own email, phone number, slack, credit card, etc.
→ models - it should use a combination of models under the hood, not just one. Different models have different capabilities.
→ memory - it should have an amazing memory.
→ reliability & consistency
→ can self heal
→ has the ability to be always on (i.e lives in its own container)
If you essentially have these parts in place, you have created a very powerful entity. You have created, in some sense, a digital employee. It can virtually do anything you need it to on a computer - i.e in theory there is nothing it can't do that you can do on the computer. Here's why this is massive.
Think about most of the work you do on a computer. Any software that isn't related to entertainment is a means to an end. Humans have needed to multiply numbers and do complex calculations to figure out many things. We used to have buildings full of women whose jobs were to do calculations for the space program. Then excel came out, and one person with excel would outperform the entire building of people. Software started eating the world, because we could encode procedural logic into a piece of metal, then speed it up and copy it infinitely. This paradigm of computation became really useful for humans to delegate computation. Furthermore, networks let us communicate at lightning fast speeds. Until AI though, we had one very specific limitation - logic had to be encoded extremely procedurally. You had to clearly define inputs and outputs, and everything in between. The analogy here is you give a driver a car and a map of a very specific route they have to take to get from point A to point B. AI changes that, because you no longer need to define everything in between - i.e don't give the driver the map, just tell him where he needs to go.
So most software was built for us to get stuff done - whether it be ordering a chair for your office on amazon, do your finances on Chase.com, search for some information on a browser. The software provider let you put in the input, then defined the 'map' to get to the output. Everything you do on your computer apart from scroll TikTok for needless dopamine hits is to fulfill some quest. That is essentially what 95% of knowledge work is, and what Microsoft made a fortune doing. They lined up cubicles with computers running MSOffice, and now every office from construction to law uses Word to draft documents, Excel to do number crunching, Powerpoint for presentations, Outlook to receive emails, etc.
But you no longer need humans actually operating this stuff. If you have a powerful agent like the one I described earlier, it can effectively do all of this for you. It can connect into your powerpoint, your excel, etc. You tell it what needs to be done, it just gets done. You want a chair from amazon, just tell your agent. It will learn your preferences and find you an amazing chair for an amazing price.
Right now, there really aren't ubiquitously powerful AI agents. Even the best in-class agents today are fallible. They deteriorate over long horizons, they are very poor at many tasks. Agents are great at doing tasks that are easily verifiable, because they can keep trying until they succeed, but many tasks that we have aren't easily verifiable. Memory is not a perfectly solved problem - knowing what to remember and what to forget is a problem even humans struggle with. Furthermore, english being the mode of communication trades extreme convenience for ambiguity. And models are seemingly blackboxes of trillions of matrix multiplications, making it hard to predict and understand how and why they make certain decisions. In summary, there's yet alot of work for us to do for both making the models better and improving the infrastructure around the models to make agents better employees.
If I had to summarize the problems of coding agents today, they would be, in order of importance - reliability & naivety, inertia, safety, cost, speed.
1) Reliability & Naivety - The agents are not consistent and are prone to degrading, going down weird rabbit holes, and simply performing in random ways. They sometimes make very dumb decisions.
2) Inertia - Clicking around the computer is actually a terrible way to do things, and thanks to cloudflare & captchas, many products and sites make life very hard for agents to operate.
3) Prompt injection - this is unsolved, although you can mitigate this by running a small model on any incoming inputs. The problem is that because LLMs are non-deterministic, technically speaking a well-crafted prompt can trick both the AI security layer and then the coding agent. As AI-pilled as I am, I would not give my agent unfiltered access to my bank account.
4) Cost - especially for frontier models, is very large. Luckily for now, the labs are discounting these costs, but even for the average user, to have an always on running agent thats actually doing work will cost atleast ~$200 a month ($100 for tokens, $100 for compute).
5) Speed - SOTA models churn out tokens at around 30 tokens/second, which is unimaginably slow. But if you use a model like SWE-1.7 by Cognition hosted on Cerebras, thats at 1000 tokens/second - a world of a difference.
So where do we go from here? What predictions can we make about the future? Here's my list.
1. The agents get much better because of the models, but not existentially threatening anytime soon.
Humans are feeding AI so much training data that they're going to get better. The reasoning chains will become better over long horizons & the decision making quality will increase.
It is unclear if there is a cap on how good the models get. ASI, RSI etc. still feel like a sci-fi movie. My prediction is - AI will get much better than it is today, but it is fundamentally blocked. AI is modeled on an abstraction of reality created by humans, not actual reality. It does get a taste of reality through verifiable tasks like coding & math (if you subscribe to the notion that the consequences of mathematics is encoded into reality), but for a large part AI does not experience reality like we do.
This is a very Sutton based argument of whether the model architecture today hints at real intelligence or complex imitation machines. My perspective is that its somewhere in the middle, but I would agree that there would be a massive unlock if we let AI experience reality for itself and continually learn, and until we get there we might be fundamentally limited.
I also think there's much more compute, algorithmic efficiencies, and data that we are yet to throw at these models, which will give us great gains.
2. The harnesses increasingly get absorbed into the model.
We're starting to see a commoditization of models,and a value accrual on the harness layer (i.e the layer on top of the model which controls where context goes in). This will keep happening, but that layer will keep moving away from the model.
A big part of how the models get better is through people using them. That behavior feeds back into the models - how to manage tools, compaction, memory, skills, etc. Think about when you have a long running chat with an agent, and your harness forces compaction at certain intervals. All that data is feeding back into the models, so they will learn better how to have longer conversations.
Overtime, the harnesses will keep getting higher level, and the models will do more and more heavy lifting. This will be similar to the Forrest Gump moment we had with workflows - the harness will get less and less invasive. We are already seeing hints of this with claude code deleting large parts of their system prompt.
3. There will be a rise of specialized agents.
The implication of (1) and (2) is that the baseline agents that the labs sell will rapidly improve. And while they get better generally, they will struggle to perform on domains where they lack data.
For eg, if you try to get an agent to book you a trip to the Maldives, you would have to hand hold it at many steps - the experience would be sub-par. That's why startups like Soar.app (travel agent) can do tremendously well. They build all the tooling, infra, and data around travel, then use a combination of the best frontier models to give people a phenomenal travel agent experience. Over time, as people use their app, they will build a data moat.
You can take a model, and build a bunch of proprietary tooling and infrastructure around it, and then let other people interact with your agent to achieve a certain goal. If your specialized agent is significantly better at a task than what they get out of the box with codex, you've created value.
You have the tailwind of intelligence on your side, so your focus becomes in getting access to the data and capabilities that elevate the baseline agents at a domain, which is why I suspect there will be an influx of successful specialized agent companies You can even take advantage of combining different models with different strengths. The levers are harness + context + tools + capabilities.
We see this with Basis AI (accounting), Tsenta (job hunting), Greptile (code review).
There are things that the model providers will largely win because the baseline agent simply solves the problem. But if the task is consequential enough, there will be a market that cares for supreme quality, and then you have to beat the model by enough of a margin for those people to care. For eg. OCR (a small margin of error is devastating).
The last variable to consider here is the market size - because if it is large enough and easy enough (on their path), the labs will also be compelled to focus some energy and try to dominate. This doesn't mean you can't win though - Devin is a great counterexample. Its just worth noting.
4. Smaller models will be used more and more.
Right now tokens of frontier models are outputted at 30 tokens/sec, and with Cerebras they will presumably be outputted 30x faster. Devins SWE 1.7 model feels magical. Rate of change of complex work is bottlenecked by token output, and this will go away.
More interestingly, the cost per unit intelligence is decreasing by a factor of 10 every year due to knowledge distillation. To be clear, I am not referring to token cost. SoTa models will probably get more expensive, but what you pay for some amount of intelligence today will get 10x cheaper.
We don't need Fable 5 for most tasks, its severely overkill. Models getting 10x cheaper is a massive tailwind for the democratization & diffusion of AI.
5. Using a computer becomes a novelty.
I know this is a hot take, but bear with me.
Almost everything you do on a computer is for a means to an end. Computing let us write programs that ran on the machine to perform millions of operations in seconds, and the internet gave us a network for these programs to be distributed. Databases let us store information. Boiling it down to these three categories - writing programs to do work, being able to share it to everyone, and storing memory - inadvertently became all we needed. So, software started eating the world. Everything we could possibly digitize, we did, because it was convenient to do so. We did have a bottleneck though - software engineers were scarce & expensive. Thats gone now.
Not only can AI agents write software, they can use computers and use this software. You can ask Fable 5 to go to the San Francisco Tennis Reservation System and book a tennis court for you without having to click through all the pages. Looking through your inbox is a thing of the past, an agent can look through it for you, summarize & manage it, and even draft messages for the important stuff.
Browser-use & computer-use are the worst way to actually use software. Its token intensive and slow - the agent has to look at the DOM (html) and the screen to figure things out. But the internet was built over 20 years, and not everyone is going to ship a shiny MCP and programmatic access for agents to come visit. You still have several websites that conditionally render .cshtml as you scroll. And worse - you have several sites that don't actually want you using AI agents to access them - so they try to block agents with heinous captchas.
The writing is on the wall that software from here-on-forward will be built with the end-users as agents, and most companies will be heavily incentivized to create the access for agents to be successful, which they can do through an mcp or programmatic access.
There will therefore be two camps.
websites / software that want to be agent friendly
the ones that don't.
Social media sites like Twitter & data-heavy sites like Zillow obviously don't want agents crawling around, so they'll do everything in their power to stop that from happening, while even a random parts supplier in Indiana will want some way for agents to be able to buy parts. The middle ground is then people that haven't caught up yet, and people that want to access the sites through agents which don't want to be accessed.
While the shift takes place, and products and software become increasingly AI friendly, browser agents and computer use also become better - so most of the digital world becomes accessible to AI agents, and we'll probably have a cat and mouse game on the sites that want to ban agents - which has been happening with deterministic bots and scrapers anyways.
Given that most of the digital world is accessible to AI agents, and this only increases over time, you as a human being will start to use your computer to do things less and less, until one day you will be doing all of your work through a very nice chat interface. (Plug: I'm building this at www.puffle.ai)
Your computer becomes a dumb pipe, and using your computer becomes a novelty. The AI agents will have their own computer which lives in the cloud, so they don't shut down when you close your laptop. You will use your device to communicate with your agent, but it won't live on your device. It will do all your work for you.
6. Software value hinges on data, infra, & networks
The only software products that are valuable will be those that have scarce data, need heinous infrastructure setup, or have network effects. A pure SaaS is cooked, because even if you were genuinely creative enough to find the problem, it will be almost instantly replicable. There's already tons of open source alternatives for many products we use today. SaaS's that have millions of folks using them, and large enterprises, are somewhat cushioned from this because they had to solve major scale issues - but even then. When software inevitably becomes near free, how can anyone justify paying any money to a managed solution. I recently saw a post by ShadCN complaining about how people replicate creative UI projects in an evening. Ultimately, businesses are about value add & value capture. I'm sure nice UI components will add value before prediction #5 becomes a certainty, but you can't capture significant value selling it because of how easily replicable it is.
7. The age of the physical world is back
Software had great valuations because of its multiplicity. You could scale up software so easily with such high margins that it became the most valuable industry. But now, moving bits is easy, and its only getting easier. Moving atoms, however, is hard. And as our digital ether becomes abundant, the physical world becomes a glaring bottleneck. You create value by giving society what it needs, what it finds hard to get, and largely that is the physical world. We will have self-improving self-maintained software that spins up in milliseconds, whether disposable or permanent.
8. Scientific Discoveries
This one is very tricky. I don't have a well thought out opinion on this, but I think at the very least AI will start to find a lot of correlations within pieces of work that we've missed. We've already seen this, with AI solving some of Erdol's unsolved problems. Whether AI is actually able to come up with completely new and novel theories (like complex numbers, the theory of relativity etc.) is almost more of a philosophical nature about the ontology of reality & knowledge.
If all innovation on the model front were to stop today, most professions that would largely get absorbed. Today - the following professions can be automated completely - customer service, any clerical work, secretaries and executive assistants, IT support, HR, supply planning, records, mail, admin, etc, which together combine ~ 5% of the GDP. AI is good enough to do these tasks now, because these tasks have well structured inputs and verifiable outputs.
For most of the population, these tasks are fairly mechanical and not as complex. Assuming the models do get better, complexity also stops being a barrier, and the higher skills operations start to get eaten in - engineering, architecture, research, design.
The top percentage of humans at their fields will still be valuable. Specifically, humans that are good at collapsing an extremely arbitrary action space into long chains of executable tasks using taste, design, and creativity, are fairly irreplaceable. As AI models get better, more and more humans fall below the rising sea level of value add.
Over the next decade, as AI diffuses into the economy, the shift will become much more evident. We're only a few years in