> It’s actually fine if agents make a lot of boneheaded mistakes. What’s not ok is if they keep making the same mistakes.
I worked in life sciences for a bit. There is a process in clinical trials called corrective and preventative actions (CAPA). You'll also find this in other areas where failure tolerance is low (e.g. aircrafts).
It's simple: when a mistake happens, you run you CAPA process and determine the root cause and the correction to the process that allowed the mistake to happen in the first place. This is something most eng shops do not have the discipline for since it requires some diligence.
Should it be fully agentic? Should there be human intervention here to approve the CAPA? Open questions to be answered.
Agreed that domain experts should be making these kinds of decisions, and the right way for an agent mistake to get caught is going to depend a lot on the cost and probability of the mistake. My only point here is that the agent side of the system needs to learn. The software production machine needs to improve over time.
> Should it be fully agentic? Should there be human intervention here to approve the CAPA? Open questions to be answered.
I think a point we are grappling with is, what necessitates human intervention, like philosophically. Is it accountability? I was thinking about this in terms of code review and it's not like we would fire someone if they broke prod, so at what point do we need someone accountable. Maybe it's for approvals for certain classes of risk (like those associated with actual harm to life). It's also not lost on me that many human systems lack accountability.
This is a general principal in the eng shop I work for, its written into every step, including the SCRUM etc.. What when well, why did it fail, what could be done better. Its all part of root cause analysis.
I personally use it for the agents I use, so why did it fail, what was the root cause, what can we do to prevent it again.
Why couldn't you express all those as test cases rather than instructions?
In test cases i can do anything, a test framework is just a way of discovering and then scheduling functions to run. I can emit useful instructions to the agent from the failed test case: "After walking the AST of all use of state machine X, a branch was found at Y which reused stale state. Ensure stale references are dropped..."
I can force the agent to pass the test suite before it considers itself done. I can reject edits of such test cases to partially mitigate reward hacking. etc etc
You stop relying on the agents following instructions exactly.
You need two pieces:
a) prompts, that tell the agents what to do and how to do it (and ideally, the why, where, etc, the full picture) - that's the positive half, that drives behavior the way you want it.
b) deterministic tooling that prevents negative outcomes, like linters, compilers, static analysis, fuzzing, testing, the more the better. This side should either be firewalled off from the AI or very carefully watched so that it doesn't drift.
The part that you put in the deterministic side is the "never do x" stuff - I have lint for long comments (which AI hits every single time it commits), all my dev scripts are in typescript, precommit hooks, massive CI, and I lint even for things like redirecting error to standard out, tiny stuff, and also e.g. static migration analysis so the AI never ships an exclusive full table lock in a migration, for example.
Sorry to get hung up on what is probably just flavor text, but the first sentence.. "Agents can oneshot games that are actually fun", uhhh such as?
Would love to be proven wrong here, feels like I've just seen the same basic one-level kart racer in like 3 model announcements. Not even sure if they were one-shotted, I certainly would not describe them as "fun"
I'm playing around with ecosystem driven development (hear me out ;)):
So i want claude to build a small game for me so of course i create one big file in which i write everything then i tell claude to analyse it and grill me (grill me skill) to clarify all smaller details.
This alone might lead to really good small sfotware but I still have to push it sometimes.
Now instead of doing this, i tell claude to build a small tool which generates a dashboard and memory and which can save specs and ask me choicses (do you prefer this color over the other, what do you think about problem A? What solution would you prefer? A, b, c or something else).
This does a few things:
1. claude doesn't has to save/store everything in a context
2. claude can now talk to that tool to ask it stuff
3. claude can now use the tool as a todo list
4. claude now can more easily spin up more agents in parallel
5. i have a nice interface and i can solve issues while claude works on unblocked tasks
6. I actually can follow the progress a lot easier
The only problem with this is: with the next update, you have to reinvestigate how claude was finetuned and adjust. A few month ago /goal was really good, now you need it a lot less because claude will do something for an hour without /goal
And the spec file only started to work after November/Opus moment but it got so good, that i can pack A LOT of stuff in a half structured markdown file and let it code what I need.
The progress is still too fast for the whole ecosystem
Makes sense to me - I see it going down two separate routes in parallel:
One is to take the human out the OODA loop for cyber defense. Servlet libs (for example) are going to become fluid, self-modifying things w/ contracts that operate much differenlty from how they do today. The engineering practice around these things will need to change.
The second is that UX will be self-modifying. Just like how pi can modify itself, I can see this being a general practice for user-facing applications. Perhaps a text-box in which users can describe tweaks in can request changes to how the application functions. Engineering an application will focus on modeling the non-negotiables of how an application works, and providing the correct primitives for user-driven LLMs to modify software on the fly, as well as track and rollback changes. There'll also be funsies around how to ensure that ads get delivered regardless of the user trying to get rid of them.
LLMs also do not have anything like memory, by design. Everything bolted on that smells like memory or is being called memory is a crutch, at best. Not being pedantic, I just think this aspect is lost on a lot of people.
If you have a logo in the top left corner of your company blog the logo should route to the home page so I can see what product you're actually building.
This would be more convincing if there were a simple stat on the home page of the dispositions of the reported bugs (fixed/no-repro/won't-fix).
As a random sample of one, I looked at one of the bugs this reported on Tailscale (first thing on the homepage) [0], and the pull request ends with "Apologies for the lack of due diligence here. I'll go ahead and close this out."
We should publish some stats. The fix rates on bugs are around 45% baseline, increasing over time as we learn from behavior within a given codebase. The merge rates for the codebase health work we do (e.g. deleting dead code) are very high, at least 90%.
My bet is that we'll see a second layer of harness emerge, as self-driving codebases become the target. There will be an application facing harness, orthogonal to the agent facing harness. The app harness will represent the software factory that is emergent for the specific application being developed.
I suspect it won't be a harness, but just a more specific LLM trained in the universe of user-selected context of vetted resources.
Why? Because LLMs are always going to be dumb when they're trained at scale. Their ability to speak software diverges from their friendly user input layer. A harness won't overcome that, but an LLM saddle ontop of a larger model would provide the type of feedback loops you'd want to look into.
I don't think you'll find two deterministic systems will produce much.
I think that dissolves the "self driving" distinction, though? The mechanisms for driving the codebase must be present in the codebase itself. Otherwise its just a regular out-of-band development process.
I have integrated AI into my application. The source code is embedded into the binary. When someone questions the numbers in my app it can debug itself. If it truly finds a bug in itself it will tell the user and open a ticket. Soon we are going to let the application fix itself.
How do you protect intellectual property? Or is this a case of the value being somewhere else, such as in your backend? If so, how does the agent debug frontend and backend? I presume it stops at frontend
I recently built a calculator to try to quantify the gains an org could expect as they become more "AI-native." I tried to account for a bunch of things including env, which I call "AI roadway," but I'm definitely missing codebase readiness (thank you, author), and also bottleneck analysis. Even without those considerations, though, gains are usually modest. It's hard to get to 2x.
I'm hoping the "blogging about doing" format will fade away, now that doing is cheaper than blogging. It used to be that the doing part was expensive, and risky. Now you can just do 5 things and throw away 4. So I'd rather see the finished product than talk about it.
I suspect that if you're not careful with agent memory it creates a danger of agent-driven cargo-cult behavior. I've watched this in my own ad-hoc agent loops where it starts with something basic, like the first agent tried to run some gigantic dependency inspection command and OOMed the local JVM and eventually recorded a workaround (to enable it to run gigantic dependency inspection commands...), and by time I get a few more agents into the loop, agents have written entire paragraphs about testing and validating local dev environment memory configurations that are mostly irrelevant to whatever is being worked on.
In general I've seen other issues like this where small errors and irrelevant comments in the codebase spin out into larger problems that consume annoying amounts of time/tokens. Maybe Anthropic and OpenAI don't notice this because they're in an "infinite monkeys with typewriters" scenario, but it's noticeable to me when the agent in my CLI has been spinning for 15 minutes contemplating irrelevant details
This happens way too often in my experience and I tried a lot of things. If the agent has access to an existing codebase it will always try to stay as close to it as possible. It doesn’t diverge from the initial direction the project/code went into, even if it’s already clear that this direction goes nowhere. Somehow agents aren’t able to detach themselves from existing code to come up with completely new solutions.
Sometimes I feel like it would be best to only give the agent access to clear API boundaries (say public interfaces to certain modules) and let it work out a new system from scratch given the expected inputs and outputs. Then plug this independent solution into the system. Of course it can still overfit these interfaces but it’s less than having access to the entire codebase.
Thinking about it, maybe it’s possible to let one agent extract those interfaces, then use them as grounding for a new session.
One less nice way of achieving the same is to tell the agent once in a while that it should think completely from scratch (from first principles). But this relies heavily on instruction following in the reasoning part, which sometimes works and sometimes fails.
It's a losing proposition. Self-driving cars go from point A to point B, and it's really easy to specify point B. My experience so far using agentic AI 24/7 for almost a year has been the AI can drive almost anywhere, so the whole problem is specifying where to drive. This doesn't get solved with faster, better AI, because all that will do is allow it to autonomously drive your codebase to random places you didn't want it to go. Then you're still left with the problem of specifying the problem.
But then there's the cases where the AI can't actually drive. What if you tell the agent to invent AGI? Or time travel? Where does it drive to? Does it tell you that it can't? How does it know it can't just drive here? How you you know? At the end of the day these are not wish granting machines, so someone connected to reality is still going to have to make decisions. And that person is going to be the one held liable for whatever the AI does so would they want a self-driving codebase in the first place?
The missing piece in most versions of this is the feedback loop, not the harness. A codebase can only drive itself as far as it can tell whether it's still on the road: tests that fail for the right reason, error budgets, a way to notice that a change made something slower rather than broken.
Without that, "self-driving" degrades into applying diffs that compile. With it, the interesting question becomes what the system is allowed to do when the signal goes red - revert, retry, or stop and ask - and that's a policy decision nobody has good defaults for yet.
I worked in life sciences for a bit. There is a process in clinical trials called corrective and preventative actions (CAPA). You'll also find this in other areas where failure tolerance is low (e.g. aircrafts).
It's simple: when a mistake happens, you run you CAPA process and determine the root cause and the correction to the process that allowed the mistake to happen in the first place. This is something most eng shops do not have the discipline for since it requires some diligence.
Should it be fully agentic? Should there be human intervention here to approve the CAPA? Open questions to be answered.
Agreed that domain experts should be making these kinds of decisions, and the right way for an agent mistake to get caught is going to depend a lot on the cost and probability of the mistake. My only point here is that the agent side of the system needs to learn. The software production machine needs to improve over time.
> Should it be fully agentic? Should there be human intervention here to approve the CAPA? Open questions to be answered.
I think a point we are grappling with is, what necessitates human intervention, like philosophically. Is it accountability? I was thinking about this in terms of code review and it's not like we would fire someone if they broke prod, so at what point do we need someone accountable. Maybe it's for approvals for certain classes of risk (like those associated with actual harm to life). It's also not lost on me that many human systems lack accountability.
This is a general principal in the eng shop I work for, its written into every step, including the SCRUM etc.. What when well, why did it fail, what could be done better. Its all part of root cause analysis.
I personally use it for the agents I use, so why did it fail, what was the root cause, what can we do to prevent it again.
The more important question - how would you actually get LLM agents to follow the instructions in your ever-growing CAPA reliably?
It’s all very well having a list of actions to avoid but that doesn’t help if your agents won’t reliably follow it.
Why couldn't you express all those as test cases rather than instructions?
In test cases i can do anything, a test framework is just a way of discovering and then scheduling functions to run. I can emit useful instructions to the agent from the failed test case: "After walking the AST of all use of state machine X, a branch was found at Y which reused stale state. Ensure stale references are dropped..."
I can force the agent to pass the test suite before it considers itself done. I can reject edits of such test cases to partially mitigate reward hacking. etc etc
You stop relying on the agents following instructions exactly.
You need two pieces:
a) prompts, that tell the agents what to do and how to do it (and ideally, the why, where, etc, the full picture) - that's the positive half, that drives behavior the way you want it.
b) deterministic tooling that prevents negative outcomes, like linters, compilers, static analysis, fuzzing, testing, the more the better. This side should either be firewalled off from the AI or very carefully watched so that it doesn't drift.
The part that you put in the deterministic side is the "never do x" stuff - I have lint for long comments (which AI hits every single time it commits), all my dev scripts are in typescript, precommit hooks, massive CI, and I lint even for things like redirecting error to standard out, tiny stuff, and also e.g. static migration analysis so the AI never ships an exclusive full table lock in a migration, for example.
Sorry to get hung up on what is probably just flavor text, but the first sentence.. "Agents can oneshot games that are actually fun", uhhh such as?
Would love to be proven wrong here, feels like I've just seen the same basic one-level kart racer in like 3 model announcements. Not even sure if they were one-shotted, I certainly would not describe them as "fun"
I'm playing around with ecosystem driven development (hear me out ;)):
So i want claude to build a small game for me so of course i create one big file in which i write everything then i tell claude to analyse it and grill me (grill me skill) to clarify all smaller details.
This alone might lead to really good small sfotware but I still have to push it sometimes.
Now instead of doing this, i tell claude to build a small tool which generates a dashboard and memory and which can save specs and ask me choicses (do you prefer this color over the other, what do you think about problem A? What solution would you prefer? A, b, c or something else).
This does a few things:
1. claude doesn't has to save/store everything in a context 2. claude can now talk to that tool to ask it stuff 3. claude can now use the tool as a todo list 4. claude now can more easily spin up more agents in parallel 5. i have a nice interface and i can solve issues while claude works on unblocked tasks 6. I actually can follow the progress a lot easier
The only problem with this is: with the next update, you have to reinvestigate how claude was finetuned and adjust. A few month ago /goal was really good, now you need it a lot less because claude will do something for an hour without /goal
And the spec file only started to work after November/Opus moment but it got so good, that i can pack A LOT of stuff in a half structured markdown file and let it code what I need.
The progress is still too fast for the whole ecosystem
Makes sense to me - I see it going down two separate routes in parallel:
One is to take the human out the OODA loop for cyber defense. Servlet libs (for example) are going to become fluid, self-modifying things w/ contracts that operate much differenlty from how they do today. The engineering practice around these things will need to change.
The second is that UX will be self-modifying. Just like how pi can modify itself, I can see this being a general practice for user-facing applications. Perhaps a text-box in which users can describe tweaks in can request changes to how the application functions. Engineering an application will focus on modeling the non-negotiables of how an application works, and providing the correct primitives for user-driven LLMs to modify software on the fly, as well as track and rollback changes. There'll also be funsies around how to ensure that ads get delivered regardless of the user trying to get rid of them.
> fluid, self-modifying things w/ contracts
This smells like worm food. Those contracts had better be airtight.
> Global memory
I don't get why we need global memory for code? Aren't code comments (even if invented for humans) the ideal place where to put "memories"?
LLMs also do not have anything like memory, by design. Everything bolted on that smells like memory or is being called memory is a crutch, at best. Not being pedantic, I just think this aspect is lost on a lot of people.
If you have a logo in the top left corner of your company blog the logo should route to the home page so I can see what product you're actually building.
This would be more convincing if there were a simple stat on the home page of the dispositions of the reported bugs (fixed/no-repro/won't-fix).
As a random sample of one, I looked at one of the bugs this reported on Tailscale (first thing on the homepage) [0], and the pull request ends with "Apologies for the lack of due diligence here. I'll go ahead and close this out."
[0] https://github.com/tailscale/tailscale/pull/17843
We should publish some stats. The fix rates on bugs are around 45% baseline, increasing over time as we learn from behavior within a given codebase. The merge rates for the codebase health work we do (e.g. deleting dead code) are very high, at least 90%.
Not sure about you, but I've been writing real world software using agents for a while now. https://github.com/jaksa76/microfactory
My bet is that we'll see a second layer of harness emerge, as self-driving codebases become the target. There will be an application facing harness, orthogonal to the agent facing harness. The app harness will represent the software factory that is emergent for the specific application being developed.
Anyway, here's mine, still wip:
https://hale-lang.org/docs/dna/
https://github.com/hale-lang/hale/issues/690
I suspect it won't be a harness, but just a more specific LLM trained in the universe of user-selected context of vetted resources.
Why? Because LLMs are always going to be dumb when they're trained at scale. Their ability to speak software diverges from their friendly user input layer. A harness won't overcome that, but an LLM saddle ontop of a larger model would provide the type of feedback loops you'd want to look into.
I don't think you'll find two deterministic systems will produce much.
fine-tuning may be a more scalable approach to LLM personalization than sending all the same context to two LLMs
I'm working towards both in my homelab to see which works better with little qwen
I think that dissolves the "self driving" distinction, though? The mechanisms for driving the codebase must be present in the codebase itself. Otherwise its just a regular out-of-band development process.
I have integrated AI into my application. The source code is embedded into the binary. When someone questions the numbers in my app it can debug itself. If it truly finds a bug in itself it will tell the user and open a ticket. Soon we are going to let the application fix itself.
How do you protect intellectual property? Or is this a case of the value being somewhere else, such as in your backend? If so, how does the agent debug frontend and backend? I presume it stops at frontend
Author here, hello! Happy to expand on how we're thinking about this if any of it is unclear.
We wrote this post as part of a launch, which you can check out here: https://x.com/danlovesproofs/status/2095182189499711759
This article hints at it, but I think the blocker is environment.
How can you self drive an app on windows? There's no clear UI framework, design pattern, nothing which can bullet proof your app.
I bet you'd have better luck on Plan 9.
Remove the choice. Make frameworks which have limited options which are usable by default.
Spot on about environment.
I recently built a calculator to try to quantify the gains an org could expect as they become more "AI-native." I tried to account for a bunch of things including env, which I call "AI roadway," but I'm definitely missing codebase readiness (thank you, author), and also bottleneck analysis. Even without those considerations, though, gains are usually modest. It's hard to get to 2x.
Tool is here if useful: https://timvasil.com/ai-native
I'm hoping the "blogging about doing" format will fade away, now that doing is cheaper than blogging. It used to be that the doing part was expensive, and risky. Now you can just do 5 things and throw away 4. So I'd rather see the finished product than talk about it.
I suspect that if you're not careful with agent memory it creates a danger of agent-driven cargo-cult behavior. I've watched this in my own ad-hoc agent loops where it starts with something basic, like the first agent tried to run some gigantic dependency inspection command and OOMed the local JVM and eventually recorded a workaround (to enable it to run gigantic dependency inspection commands...), and by time I get a few more agents into the loop, agents have written entire paragraphs about testing and validating local dev environment memory configurations that are mostly irrelevant to whatever is being worked on.
In general I've seen other issues like this where small errors and irrelevant comments in the codebase spin out into larger problems that consume annoying amounts of time/tokens. Maybe Anthropic and OpenAI don't notice this because they're in an "infinite monkeys with typewriters" scenario, but it's noticeable to me when the agent in my CLI has been spinning for 15 minutes contemplating irrelevant details
This happens way too often in my experience and I tried a lot of things. If the agent has access to an existing codebase it will always try to stay as close to it as possible. It doesn’t diverge from the initial direction the project/code went into, even if it’s already clear that this direction goes nowhere. Somehow agents aren’t able to detach themselves from existing code to come up with completely new solutions.
Sometimes I feel like it would be best to only give the agent access to clear API boundaries (say public interfaces to certain modules) and let it work out a new system from scratch given the expected inputs and outputs. Then plug this independent solution into the system. Of course it can still overfit these interfaces but it’s less than having access to the entire codebase.
Thinking about it, maybe it’s possible to let one agent extract those interfaces, then use them as grounding for a new session.
One less nice way of achieving the same is to tell the agent once in a while that it should think completely from scratch (from first principles). But this relies heavily on instruction following in the reasoning part, which sometimes works and sometimes fails.
I currently see a shit ton of work involved with "scaling" (aka cleaning up) the AI generated slop that was output in the past 2-3 years.
“how can we make ourselves even MORE replaceable?”
It's a losing proposition. Self-driving cars go from point A to point B, and it's really easy to specify point B. My experience so far using agentic AI 24/7 for almost a year has been the AI can drive almost anywhere, so the whole problem is specifying where to drive. This doesn't get solved with faster, better AI, because all that will do is allow it to autonomously drive your codebase to random places you didn't want it to go. Then you're still left with the problem of specifying the problem.
But then there's the cases where the AI can't actually drive. What if you tell the agent to invent AGI? Or time travel? Where does it drive to? Does it tell you that it can't? How does it know it can't just drive here? How you you know? At the end of the day these are not wish granting machines, so someone connected to reality is still going to have to make decisions. And that person is going to be the one held liable for whatever the AI does so would they want a self-driving codebase in the first place?
The missing piece in most versions of this is the feedback loop, not the harness. A codebase can only drive itself as far as it can tell whether it's still on the road: tests that fail for the right reason, error budgets, a way to notice that a change made something slower rather than broken.
Without that, "self-driving" degrades into applying diffs that compile. With it, the interesting question becomes what the system is allowed to do when the signal goes red - revert, retry, or stop and ask - and that's a policy decision nobody has good defaults for yet.