My general principal is the cost of deleting code should be as low as possible and that includes the entire application as long as the data is around and easy to repurpose then the program itself should be as deletable as possible
It lacks a critical self, but the weights are there for in context learning and nudging behaviour. It's goal is to complete whatever task is given. You need to make sure the outcome you want is clearly defined.
You don't need elaborate prompts, just a few lines
"All code must have corresponding tests written ahead of time to prove the code meets the specification" is sufficient for most use cases. Prose can help nudge it more if it isn't adhearing consistently.
The better way to do this is give them each an inbox and outbox folder and a user account on a Unix machine.
They can write to their outbox, have it be sent to the other's inbox. A watcher can watch for new messages in the inbox and trigger the other agent to read it. You can have the thread Id of the messages correspond to the session id.
We don't need to reinvent things that Unix provides.
This website and article promote the destruction of property. If you disagree with something, you can engage civily, encourage people to vote with you, run for elections. Violence is not the answer.
Hi. The mayor of Denver pushed through flock cameras despite them being unpopular and not even getting enough votes to buy them. He got them to change the price enough that he didn't need the votes to get them installed.
How do you have a civil society when the people in power cheat?
It sounds like he worked within the legal constraints of the system he was elected to work within.
This kind of discretionary spending authority can used for things that are good, bad, or indifferent. When it gets used to cut through the red tape and buy a new swingset for a neighborhood park, then that's good; nobody complains about that. (Except someone would surely complain about that, but come on man.)
And when it gets used to install government tracking systems, that's bad.
> How do you have a civil society when the people in power cheat?
The problem isn't that the mayor can spend some money. Rather, the problem here is that government tracking systems are completely legal to buy.
The laws need adjusted so that government tracking systems are completely illegal, instead.
"Yeah, good luck getting the government to do that!"
The people of Colorado are free to initiate their own legislation and constitutional amendments and then vote them into force.
Destroying a camera isn't violence. It's destruction of property, sure. But property isn't inherently good and sometimes it degrades society.
If some goober installs massive floodlights that blast into windows of some houses, I think everyone would support a kid with a slingshot busting a few bulbs. If some guy is blasting music from a speaker at 3 AM every single day, I don't think anyone will complain about a cable being cut. If cameras are installed that sell data to companies like Palantir, companies that say they want to kill you and they're going to kill you and it's just a matter of time until they kill you, destroying those cameras is the non-violent option.
"Violence" is a word normally used when the victim is sentient, but I'll go along with it:
Violence against inanimate objects is morally neutral. Violence against instruments of violence is self-defense. Violence against oppression is how the USA was founded.
A corporation has unfair political advantages including a deep purse, an unlimited lifespan, and more recently all the rights of personhood. The only advantage the people have is their numbers, and yeah numbers of votes would be great, I agree, but when votes are ignored, or never solicited in the first place, it often comes down to numbers of pitchforks, as it were.
On voting harder, see the lead incident mentioned: "This happened weeks after the city council voted to keep the cameras despite overwhelming public opposition." I also advocate patiently working through the process, but people are not blind to the trends: the democratic process is failing as government increasingly sidelines voters and the richest have the levers of power.
I don't see anything on the site or article that promote the destruction of property. It's an aggregation of public information regarding the history of vandalism towards a specific target.
The website largely documents the current state of privacy and provides resources for (digital) services that help maintain privacy. This is an encouraging civil engagement which educates and empowers the audience.
Everything I've read and learned in my 50 or so years on this planet leads me to believe that the times injustice can be corrected purely by civil engagement and voting are massively outweighed by the times that they can't. So depending on how bad the thing is - people make choices.
Without having thought about it for more than about 10 seconds: I guess I associate violence with something more personal: an actual person or living thing, or personal property. I guess "corporate property" is where it gets more into the grey zone for me.
But I see your point. Destroying a thing (even corporate) is a pretty extreme reaction that I can only see making sense after having exhausted all other "peaceable" avenues.
People that see these things as detrimental to society though are likely pretty motivated.
> FOSTA-SESTA is the source of this. A well intentioned bill, that, once again, has unintended consequences beyond it's original intention.
You're right that these are connected to FOSTA/SESTA, but you're missing the actual connection.
FOSTA/SESTA were not "well-intentioned". They were the product of lobbying from explicitly religious, anti-sex, anti-pornography groups. Those same groups are behind recent campaigns to require providing government ID to access pornography, to allow attorneys general to prosecute LGBTQ content, and to ban pornography from platforms like Steam and Itch.io.
FOSTA/SESTA have worked exactly as they were intended to! The intention was to make it harder to conduct sex work legally and safely, and they accomplished that goal!
These policies have little to do with FOSTA/SESTA themselves, in that the text of those laws has no bearing here. But those bills were the first big, national victory of these campaigns, and they used that momentum to raise absurd amounts of money to lobby for the other laws mentioned above, and to target financial infrastructure as an easy point of leverage to accomplish their goal of banning pornography across the Internet.
> Those bills are the reasons banks/credit card companies are pushing this since it holds them liable.
You are either misreading those bills or confusing them with other similar bills which did target banking infrastructure (and which thankfully did not pass).
FOSTA/SESTA did increase liability for platforms, but the applicability of those laws to this specific case is minimal to nonexistent.
Banks didn't do this before 2018 as aggressively. Reputational stuff is always been there and chargebacks have always been there. This law is why they've been so aggressive over the last few years. There's nothing else that has changed
> This law is why they've been so aggressive over the last few years. There's nothing else that has changed
There is so much that has changed! If you think FOSTA/SESTA are the only thing that have changed, you're clearly not up to date on this topic!
> Banks didn't do this before 2018 as aggressively
Because after FOSTA/SESTA passed in 2018, the groups that lobbied for it started targeting financial infrastructure as the front in their war. This is not some secret; they've been very open about it and their lobbying efforts have been extremely well-documented.
The best way to prompt an LLM is to describe the outcome you want, that's it. They are trained as task completers. A clear outcome is way better than a process.
If the LLM fails, either you didn't describe your outcome sufficiently or is misinterpreted what you said or it couldn't do it (rare).
Common errors should be encoded as context for future similar tasks, don't bloat skills with stuff that isn't shown to be necessary.
> The best way to prompt an LLM is to describe the outcome you want, that's it. They are trained as task completers. A clear outcome is way better than a process.
This is not true for anything complex. They’re instruction followers, of which task completion is just one facet.
They’re also extremely eager to complete tasks without enough information, and do it wrongly. In the case of just describing task completion, despite your best efforts, there are always some oversights or things you didn’t even realize were underspecified.
So it helps a lot to add some process around it, eg “look up relevant project conventions and information. think through how to complete the task. ask me clarifying questions to resolve ambiguities. blah blah”. This type of prompt will also help with the new Opus 4.7 adaptive thinking to ensure it thinks through the task properly.
Agreed, and further, I'd argue the OP's division of LLM instructions into either process or outcome specification is a false dichotomy. My agentic process specification is about automatically specifying the outcomes that I would otherwise repeatedly have to tell the LLM to consider, like making sure test coverage is maintained, or that decisions are documented on the original Github issue. Or it's about correcting common failure modes, like when the agent spends an enormous amount of time running repo-wide tests while debugging a focused change, because the agent doesn't consistently optimize around the time-to-implement as an outcome. Arguably part of addressing those failure modes boils down to pure process in the sense that I specify a logical order for achieving the outcomes, e.g. creating a plan before implementing. But that is mostly to organize approval gates for my convenience, rather than structuring the agent's work per se.
If there is anything we have learned in decades of Software engineering, it's "A clear outcome" is not easy to describe. In many cases, it's impossible unless people from 4 different domains collaborate. That's why process matters. It allows for software to be built is a "semi-standardized" way that can allow iterations to get us closed towards the expected outcome, that might emerge over time.
Yes, not everything I use LLMs for going to have the same level of ambiguity or complex requirements. Optimizing by choosing to skip over parts of the process is exactly Addy is talking in this article.
This seems like common sense but it does not work in practice.
Prompting is just the first part. To get the outcome, you need to have other systems to steer the agent as it get things wrong. Proper deterministic tests work. But there is also stuff that need to happen during the LLM execution like cyclic detection etc. All of this adds up.
You cannot just prompt an LLM an hope for a good outcome. It might work in small isolated scenarios but it just does not work consistently enough to call it reliable.
Without further guardrails enforced by the process or the harness, LLMs do not have sufficient capabilities to complete a task up to a certain standard.
I agree that many skills are overblown and unnecessary. But there's a lot of value in giving AI the right process. See how much more effective Claude can be for moderate or large changes when using the superpowers skill.
a skill is just reusuable/shareable context. It's just text, really. It's useful for things like documentation on how to use an API (this works better than MCP in my opinion), or a non consensus way of doing something. For example, you can use remotion to generate video. There are useful remotion skills that allow you to reliably generate specific types of videos. Captions of a certain style, for example.
That seems a bit reductive. Even with humans, there’s a range of interpretations and ways that something can be built or a task completed. Engineers remember stuff so you don’t have to keep repeating yourself. Skills are a way to describe your outcome without similar repetition.
MCPs are basically just JSON-rpc. The benefit is that if you have applications that require an API key, you can build a server to control access (especially for enterprise). It's the same as REST apis, except by following a specific convention we can take advantage of generic tools (like the one I built) and means you don't need to rely on poor documentations to connect or train a model to use your very specific CLI.
But if you have customer facing APIs then all of these problems were already solved in an enterprise context. You can force an oauth flow from skills if you want.
I don’t think that CLIs are the path forward either, but you certainly don’t have to teach a model how to use them. We’ve made internal CLIs that adhere to no best practices and expose limited docs. Models since 4o have used them with no issue.
The amount of terminal bench data is just much higher and more predictable in rl environments. Getting a non thinking model to use an MCP server, even hosted products, is an exercise in frustration compared to exposing a cli.
A lot of our work is over voice, and I’ve found zero MCPs that I haven’t immediately wanted to wrap in a tool. I’ve actually had zero MCPs perform at all (most recently last week with a dwh MCP and opus 4.6, where even the easiest queries did not work at all).
LLMs don't care about mcp vs CLI. CLIs enable LLMs to fetch/mutate data and build scripts with the same program. I think of it like a Linux dev in a box. Sometimes you want to just call a tool, sometimes you want to write a small program that calls that tool instead.
How many kernel devs does the world need? A dozen or two?
It will be the same with software. AI will be writing and consuming most software. We will be utilizing experiences built on top of that, probably generated in real time for hyper personalization. Every app on your phone will be replaced by one app. (Except maybe games, at least for a short while longer).
Everyone's treating writing code as this reverent thing. No one wrote code 100 years ago. Very few today write assembly. It will become lost because the economic neccesity is gone.
It's the end of an era, but also the beginning of a new one. Building agentic systems is really hard, a hard enough problem that we need a ton of people building those systems. AI hardware devices have barely been registered, we need engineers who can build and integrate all sorts of systems.
Engineering as a discipline will be the last job to be automated, since who do you think is going to build all the worlds automation?
How wildly dismissive of the foundation of the X$ billion dollar software industry. You think humans just stumbled into writing code by accident or something?
How does building agentic systems, a "really hard" problem, not just end up a "regular code" problem? Because that is what it is. A distributed systems problem with non-deterministic run lengths. How do you switch agent contexts? Similar to how you solve regular program context switching. How do you search tool capabilities and verify them? How do you effectively manage scheduled tasks?
Oh, look, you've just invented the operating system kernel. Suddenly, those 'dozen or two' experts don't seem so archaic after all!
Does it even make sense to build everything on top of machines that are 70% reliable? The sheer orchestration and validation overhead at scale risks being more expensive than just keeping most software engineers and having them manage a few AI agents.
Also, 200 years ago we didn't have bike mechanics. Car mechanics. Boat mechanics. Plumbers. Electricians. Not all new professions fade away.
Every problem you described is solvable and while it may not be solved right now or even in 6 months it'll probably be solved within 18 months. It's just scaling and tuning the models
You can’t “tune models” to get people willing to get on a zoom call with an agent and the agent asks them questions and talk through strategy and understand human emotions.
Are they also going to interact with the model for a design review session?
Tell the model where it got it wrong and the model is going to make the changes?
In 18 months AI agents will be able to accurately infer people's emotional state from the subtle facial expressions they make in a sales meeting, in real time?
