Using Software Engineering agents

Did you use SWE agents?
If so, which agent and model are your favourites?
I found that GPT 5.2 Codex is much better than Claude Opus 4.5, but they work quite well together in turns.

Due to the region I’m in, most of those SWE agents aren’t available where I live.

I did try GitHub Copilot in VS Code a long time ago (GPT / Claude etc). It was pretty dumb — bad suggestions, no real understanding of context. I ended up disabling Copilot completely.

I see.

Yes, at this moment I can recommend only GPT 5.2 (if you use Copilot), or Claude Opus 4.5, as all other models are pretty dumb, I’d say.

I personally use GPT 5.2 Codex xhigh (via Codex CLI) and Claude Code (with Claude Opus 4.5) as my backup, as GPT 5.2 shows the best results for me.

But I am paying quite a lot for ChatGPT Pro and Claude Max 100 subscriptions, so I can work in parallel on multiple big projects, where GPT and Claude recheck each other.

I’ve been using the new Gemini 3 with Crystal, but I’m mostly back and forth chat with it as I need to explore fixes or new features. I’m not letting it use an agent since I don’t want to deal with token exhaustion. For sure is a lot better than 2.5. Only been running into a few things it keeps wanting to do. Though it sucks on macros, but I would expect that.

I will tell you a truth: I built my LLM protocol, that helps models with proper reasoning, because I was tired of Gemini sycophancy and BS. It helped a little to Gemini, but surprisingly it was much better for Claude Opus, but GPT was the best achiever.

I am having a lot of fun with GLM from z.ai and it’s super cheap.

I am building my own SWE agent, and I use Grok 4.1 Fast reasoning that is also super cheap ($0.20/$0.05/$0.50), but quite powerful.

I think things changed quickly since this post was created. Nowadays, agents are extremely powerful compared to check 10 months ago :frowning_with_open_mouth:

Since it might be interesting, my stack is:

  • Mostly GLM 5.3 Flash at execution and simple planning, GLM 5.3 for complex planning. Into Zed Agent (Keep an eye on the project). Sometime DSV4.1 flash too.
  • Qwen 3.8 Flash Next locally run on Ubuntu with OpenCode + Playwright MCP for e2e QA; run long horizon QA, testing everything using vision + snapshot, and is run local so mostly “free”. It’s a server but with desktop package; I run OC via SSH (no display). Also used for security defect; asking the agent to try to hack the software during the night, and come back with runnable exploit. Did found a few escalation pathway in the app I am building :sweat_smile:
  • Mistral 4 for i18n tasks. Robust
  • Opus 5.5 for brainstorming, vision plans, design mockup, marketing… via the claude chat window.

Total spent:

=> OpenRouter credit, maybe $20 a month, on/off on demand. Use it also in-app, since the app I’m building has an AI assistant.
=> NeuralWatt subscription, $100 a month for like 8 billions token; most value lies here.
=> Claude 20$ subscription. Rarely run out, since I use it on/off.

Hello. 0/

I am a very happy Gemini user (Antigravity CLI). I used to have the Google One pro, which gives you some tokens. Learned to use it, created my own skills, rules and GEMINI.md (akin to a system prompt of sorts) and got very well acquainted with how it works.

Then, my brother decided to pay for Ultra and put me in his family group. Now, I have practically unlimited tokens across many projects. I am always working on 4 or 5 projects at the same time, not sloppily, though It’s me who is saying this so it qualifies as a trust-me-bro claim. ;D

In any case, I’ve put huge effort in providing “tools” for my agent like the crystal mpc server I published in gitlab, the crystal-for-agents project and skill, etc. Also, in figuring out how to work with upstream projects by developing techniques, rules, incorporating feedback and trying to adapt to projects as best as I can to cause the least friction possible, yet, with the fullest transparency about me using AI for stuff.

I would not change to any other offering even if you gave it to me. I’ve gotten to know it’s shortcomings and super powers. Hell, I could even say I’ve grown to like the AI. I have a good relationship with Gemini.

That said, I do use others, like deepseek 4.1 flash and such cheap models to gain perspective, experience and knowledge about these. I do minor tasks, experiments, multi-agent setups and such crazy things with them.

I’ve already mentioned in another topic, but here it is:

I’m very happy with this setup and I don’t feel any particular need for a better LLM model nor harness, I’m looking foward for that day that I’ll be able to run glm-5.3-flash inference locally on my workstation without having to pay for a (cheap, but still…) subscription for that.