We do not need more AI agents. We need more control over them.
That is why I created Invoker, an open-core execution engine for AI-driven engineering workflows.
The bottleneck is no longer just code generation. It is execution control.
AI work needs isolation, replay, auditability, recovery, and human decision points. The model resembles build systems and workflow engines more than it resembles a theoretically “AI-first” chat interface.
https://t.co/JkWx9kdQMH
1) Bazel for remote execution and caching
2) shift heavy tests like e2e into a nightly and have AI automatically submit repairs
3) smarter sharding
4) hyperoptimizing each part of the build step by having AI focus on it. Requires wiring your own evals infra for it but worth it.
If you want to be EXTREMELY fancy?
Try a Content addressable file system.
Android's source code is huge. 1k git repos, 1.62 million files, and it takes 300GB for a built workspace.
How did we reduce checkout time from 36 minutes to 8 seconds? Full builds from 80 minutes to 2 minutes?
We just open sourced the solution.
https://t.co/8lcqRmd5FI
i learnt this lesson this year. it's helped me follow through on my ideas and not get stuck overthinking it
don't ask if you should do the thing. do it, then come back with a demo or something tangible
you can just do things
@Siddhant_K_code I even think the metrics are suspect since all the models are trained on them and the model providers select the benchmarks
I haven’t seen a real metric saying “we outperform Claude on these exact sessions and you spend $x”
hot take: you don't need a software factory, and i don't love the term. it's just marketing
all you need is good software engineering and investing your tokens in the right places. the more deterministic the better (and token efficient)
@arunkumar_9t2 Imagine making push instead of polling a feature and calling it a win.
We’re taking SWE best practices and Leetcode answers and calling it an AI improvement.
What are we doing here 😂
I completely disagree with your reasoning but directionally agree with your conclusion
If you look at what AI is really used for and the definition of “get things done”, it becomes
1. Toil tasks. These are highly automatable (filing task tickets, sending emails) and where the heavy benefits of AI are. That alone recovers 1-2 out of an employees week. Well worth the price.
But this is not all tasks. And it’s a mistake to believe this time is reallocated to slacking.
Nobody’s raison d’etre is sending coordination emails or PM tickets.
2. Planning and testing. This is probably better. Spike and validate, experience, research. You can run hundreds of experiments to validate information and come up with a thesis.
For the most part, this is also worth the price. The problem is that this can become prohibitively expensive when agents run the wrong way.
3. Execution. Most people blow through AI budgets here by executing precisely the wrong thing for long execution times.
Planning is supposed to eliminate a large part of it but there is very little premium placed on it (because everyone is screaming about AI execution and efficiency) and completely ignoring its benefits.
But more importantly AI is being used as an odd combination of papering over the neglected infrastructure and processes companies have wrought.
If you read through all of what I said before you’ll instantly recognize what I mean.
Jira tickets? AI is being used to fulfill the coordination tax of organizations
Testing and validation. If code and information is opaque and difficult to read and understand or is flaky and untrustworthy , you get really extremely bad proofs and shit execution.
And this is where the real problems are. Companies are not seeing AI benefits because they’ve neglected the infrastructure of their operations. Operations are currently oriented towards complexity and obfuscation for job and organizational security (reasons far too long for this tweet). And layoffs makes this worse.
Giving everyone AI is like giving everyone a sports car but makign them drive on dirt unpaved roads. It s perpetual tax that gets worse with speed.
So you’re probably asking why not invest in improving the infrastructure to move faster? That’s not how companies work. Infrastructure and operations work generally suffers from lack of legibility. Products that print revenue for the company get prioritized. Maintenance work gets disrespected.
And this is why you aren’t necessarily seeing the benefits of AI. AI success will come to those who can acknowledge they have an issue that starts with the company and its polices.
Getting companies to self-admit they have problems is not going to happen.
Only The Paranoid Survive is a warning from the exception that escaped and were able to admit that.
1/16 the ai industry is making an unusual bet: buy compute for years, sell intelligence by the token. banks, commodity merchants, airlines, utilities and cloud providers each explain a different part of the risk.
These Anthropic blogs are making me realize we’re basically learning MMORPGs from first principles
They’re actually suggesting to NOT send your main unit into the (informational) fog of war without scouting and verifying with disposable units/subagents
@suchenzang@TorteDeLini should have been given equity 😹.

I remember cheesing the bots at the Redbull booth by creep cutting into the jungle.
Also OAI beat a pro team the next year iirc
you can laugh at this, but I'm pretty sure that if Dendi had won, OpenAI would've raised far less money afterward and AI progress would've been slower
back then, OpenAI was a relatively unknown research lab (~50 people) and this was the first time (after the initial launch) OpenAI made it to the mainstream news