Agentic Development Updates - October 2026
The LLM (AI) hype became disruptive. While frontier labs as Claude and OpenAI has been racing for providing better harnesses an entire market of open source harnesses grew from up. Now most of the provider and LLM aggregators (providers who resells you access to LLM) are building own harnesses as well.
This race gives us some clues:
- There are leaders in the race, there are no winners in the race
- Providers are fighting for the attention. It's not enough to provide access to models. A model must be surrounded with a suitable pipeline.
- Proprietary solutions might provide better results out of the box for a potentially higher costs, caused by the "secret sauce" - prompts, tool calling, loops, caching policies.
Understanding these mechanics lead me to a conclusion: wanna win trubines - choose a car that can be maintained and rebuilt. In simple words: do not rely on promisses and ready solutions, instead learn to maintain and adjust the harness.
This article will be full of names, providers and tools more or less known. I'm not affiliated with any of these, all the experiments has been done in my own time and paid with my or my companie's money. I explicitly intend to share and show there is a huge choice of tool, so you don't have to stick to the "this month's best" and later regret your decision being "behind" for a few upcoming months.
Where it started
A year ago I started with Claude Codex. A CLI dedicated for Software Developers and other professionals working strictly with the text files, mostly code and configs. October 2025 - April 2026 was bringing improvements to their models and results were becoming better and better.
In February I installed OpenClaw and figured out Anthropic are not happy with us using regular subscriptions to power 3rd party harness. This was a clear sign they tend to keep entire system closed and tied together.
After I've been struggling to understand limit policies under 100$ plan in April I quit the Antropic train.
Altenratives were on their way
At that time Codex became a thing with a few useful features. OpenAI is accompanying me for a few years already, from around GPT-3.5 and switching to Codex was natural. Yet I didn't like the way everything was about to be mixed together. Months later Codex was left as Codex CLI but everything else became a ChatGPT app.
By the way if you didn't know - older plain chat experience can be kept as ChatGPT Classis and it is still the most used AI assistant. It get all the questions and discussion and let me easily separate thinking and learning sessions from agentic sessions.
But Open Weight models were chasing the biggest one and I started to search for a provider. This way I found Open Code - the harness and their Go plan. By the way they have released v2 version recently - a major version that's been reworked deeply.
But the tokenomic has no mercy. And prices are going up. That's how I went to CommandCode - being a directl competitor of OpenCode with a larger model offer by that time. I'm still staying with thm and their GOAT plan.
But the harness? I went with Twigpine(ex. Gitlawb)/zero written in Go. This sounded like a perfect solution for me, as I love Go and I enjoy working with their asynchronous model. I have made some arly contributions to add OpenCode as a pre-defined provider and some adjustments, untill I hit a wall - AI reviewed PRs. I decided I'm not gonna waste my time nor my tokens to deal with that.
There is no winner and the king, keep adapting
And now I'm switching to try "pi" - the lightweight coding harness. This time paired with freshly released OCR - Open Code Review by Alibaba. I know, OCR is very misleading name, but it is what it is. And the reason for that is I found zero being unstable and randomly getting into stale state. For autonomously running harness, often controlled by another harness (openclaw) that's not the way to go.
And this time I learned it's easier to switch the app, rather than dig through AIed codebase. Since I decided not to waste my time nor tokens - I stay with it. Building with AI is one thing. It taught me an important lesson: building with AI is one thing. Building with the AI and the community - is completely different scenario. As a leader I will be very carefull gatekeeping anything with the AI, as initiative must not be borrowed in AI mess, unlike our CVs these days.
First impression on pi is very positive - responsive, fast, not lagging or going stale. Commandcode available via plugin. OCR skill is picked from a generic ".agents/skills". Sounds great for a drop-in replacement to any common coding harness. And if that won't work - I'll try the refrehsed OpenCode.
Summary
There are many flavours and we could play all day long with harnesses, polishing it or even building our own. But even with the agents and LLMs it still creates maintenance costs. It seems like we are arraving at a place where we have a choice again. But in order to make those choices - we need to get a few levels deeper, as usual. There's still a thing to remember: we have to meet people at least halfway. If someone wants to talk - go and do it. No amount of features can have more value than your users, contributers and team.
