Leonid Bugaev

Building Proof to reduce the work needed to trust a software change. Engineer, founder. I want fewer compromises between shipping and quality.

[Leonid Bugaev @buger]

Have become top 3 trending developer on Github, with @mitchellh right behind me 😱

[Leonid Bugaev @buger]

@marclou @malika_builds Or imagine committing to buying subscription from everyone who finished, and attached the billing. Even if it is partly fake - getting your first $ to your own service is an amazing feeling 😀

[Leonid Bugaev @buger]

@marclou Does your Paddle integration support filtering verified revenue and MRR by specific Paddle product IDs? I have a few independent businesses using one Paddle Billing account and want to list them separately without combining their revenue?

[Leonid Bugaev @buger]

@marclou How you would validate it was started on Nov 16? Needs github source code as proof?
(would love to try!)

[Leonid Bugaev @buger]

@imantiai Link?

[Leonid Bugaev @buger]

Amazing!. My main question is: if the code already has the wrong assumptions and some bugs inside of it, the formalized version will have the same bugs as well. Do you handle those cases somehow? Do you read the original commit messages, pull requests, or maybe task management system tickets to extract some real intent: what this functionality is supposed to do versus how it's done right now?

[Leonid Bugaev @buger]

@mikesulsenti Agree, and have similar measurements. It’s too overfitted on benchmarks.

[Leonid Bugaev @buger]

@Shpigford @conductor_build Orca?

[Leonid Bugaev @buger]

I don't think it's about the diff at all. I think the main challenge in the new agentic era is how we can compress evidence to the bare minimum level without sacrificing the quality. You will be able to review it for 20, 30 minutes without looking manually at the code, without testing it manually.

You can check at the format which my agents are doing https://t.co/2WSTxLOVxj to get some inspiration, check the attached image too.

However, to get the evidence on such a level you also need extremely high-quality engineering, which follows all of the checklists. In my case, it also means building the semantic graph of what your product does, then comparing this graph with what is on the pull request, seeing all of the differences, and flagging them (especially if the agent went out of its original boundaries). It's also about doing the full hazard analysis: what are the worst-case scenarios that could happen with this change, and how was it proven that it is safe? Have you tested all the major golden flows for the users? Do you have a video before, after, fix, and so on?

In the end we just wanna trust this change, green CI is just one of the checks.

[Leonid Bugaev @buger]

@Clevernesspc1 Yep! At the end of months will have exact data, and I for sure will utilise 100% out of it.

[Leonid Bugaev @buger]

Based on the first 3 days of Grok usage on fresh billing cycle - it now has 2x less tokens for Grok Bot usage.

In fact it is more like 4x less then old Ultra plan, because previous plan had 2 subscriptions Cursor and Grok one, now it is single.

Will see what end of the month will show, but Grok bot started eating usage like crazy, so defo looking at multi-vendor alternative for it.

[Leonid Bugaev @buger]

https://t.co/XMP4iRwk9t

[Leonid Bugaev @buger]

Your MCP's training plan:

Pick one job.
Write three ways to ask for it.
Define what success means.
Run them in two agent apps.
Read every tool call.
Fix the first gap. Rerun.

That's the idea behind Spirit Room: a Hyperbolic Time Chamber for your tools.

https://t.co/uw6FkSU5qU

[Leonid Bugaev @buger]

Does everyone have this crazy Singapore traffic spikes in their web analytics?

[Leonid Bugaev @buger]

https://t.co/XMP4iRwk9t

[Leonid Bugaev @buger]

I feel that the Codex team has slowed down the model responses at least two times after they decided to roll back use limits for the pro users. Even Sol 6.1 feels painfully slow now.

[Leonid Bugaev @buger]

The same issue with a lot of the security patches. The rsync CVE release was waiting more than a month to get the fix (not sure if it finally landed) and justification from the team (after weeks of chasing them) was that they "felt" that such a big patch can introduce even more bugs, so they would rather wait and see what happens...

[Leonid Bugaev @buger]

I recently had a meeting with our Google Vertex account managers. The first thing that struck me is that they are trying to fight for me against another Google department, trying to convert me from using AI Studio to Vertex AI. Then they try to offer a 10% discount if I commit to some predictable usage, which sounded okay.

I went to verify how much it will save for us and checked the prices of the current models. What I have found is that starting from January 1 there will be a 100% price increase for models like Gemini 3.8 Flash. 100% Carl!

I really liked that model for workflows, and company brain cases. SLA was great - just amazing workhorse. But would it worth with this hike?

To be absolutely open we were looking for alternatives for quite a while and were trying to migrate off to OSS models like GLM 5.3, which is in a similar price/intelligence category - and it was way below in terms of intelligence and speed (yet I have not fine tuned prompts and harness for it).

Seems like it is a time for aggressive optimisations, actually measuring every use-case, tuning prompts and hand picking models for it!

[Leonid Bugaev @buger]

@hunvreus Not talking that they going to 2x prices starting from Jan 😥

[Leonid Bugaev @buger]

Does everyone have this crazy Singapore traffic spikes in their web analytics?

[Leonid Bugaev @buger]

@tibo_maker https://t.co/MFu1unz8k6 - force agents to do such quality engineering, and evidence for its changes, that even the most stubborn QA will approve it without doing manual testing.

[Leonid Bugaev @buger]

@AustinAAmato Partial pass is also the data. Every learning is very valuable.

[Leonid Bugaev @buger]

What I'm building: Spirit Room, a test environment for MCP workflows.

Start with a user job and an expected result. Run it across agent apps. Inspect where it breaks. Test the fix.

Currently a hands-on pilot. What job should your MCP pass reliably?
https://t.co/uw6FkSTxBm