Well, I'd love to use an open-source harness! My issue is that Opus 5.5 / Fable 5.1 has been the best model for me to work with, and it's way cheaper to use it with the subscription to Claude Code Max. I won't pay more just to use an open-source harness. I'd rather just wait it out and stay on Claude Code not having to constantly jump around / switch tools as the industry matures.
Pi can utilize Claude or GPT subscriptions and they work wonderfully. You can have both subscriptions logged in and write your own tools to route to the model that works best in every scenario. I have access to both pro level subscriptions for work and a large credit count for api calls for Claude. I use all 3 with JEV as my router with a local model for context management and compaction. my token usage has gone down to a point where I rarely even use api calls. I would recommend oh-mi-pi for anyone who wants to get started with using an open source harness as it has very good memory and context management tools that ship with it and are extremely easy to set up.
You can use your claude subscription just fine in open harnesses!
I make one (https://github.com/juggler-ai/juggler) and it'll auto-detect and use your claude CLI. I use that claude plan for 90% of my work, plus a bit of GPT, I just prefer to do it inside a nicer, less locked-in environment than claude.
There's really no moat around proprietary closed harnesses, it's trivial to jump between them, but people need to be aware that it's possible!
Locked in? I routinely switch between claude and codex, same codebases, same skills, and even shared "memories" with maybe 5 minute effort to set it up. There is no lock in whatsoever.
The big hoo-har about it earlier this year was because something (opencode?) was pretending to be the claude CLI and using the API that way. Anthropic made a fuss, then GPT said "we don't mind people using their own harnesses" and everyone relaxed about it, as long as the harness goes via the claude CLI and doesn't try to cheat.
Yeah there's even an "Agent SDK" from Anthropic that manages Claude CLI sessions for you, and a separate token budget for apps that use it, on top of the subscription tokens.
even that was relaxed by Anthropic and they stopped chasing token prices for Agent SDK. I routinely use Agent SDK with my subscription and it doesn't come out of a separate budget. Their ToS vaguely says just don't use your personal subscription for commercial resale purposes.
You don't have to pay more - I wouldn't switch to an open source harness on a whim, but your harness/cheaper agents can always call Claude Code headlessly.
As context: I work as a consultant and produce code on behalf of my customers. What's most important to me is that the quality of the code is as high as it can be, without me having to rerun it with multiple different models to see which agents produces the best. If Opus 5.5/Fable 5.1 (or a different model for that sake) consistently produces the best code, I will use that for all my tasks, since that lowers the cognitive burden.
The cheaper agents never really matter, because the Max subscription costs me $100 and that is less than an hour of pay. In some cases the customer even pays for the subscription.
But I do love the idea of being able to use Pi or OpenCode. I am a heavy terminal user, and used Vim/Tmux/Lazygit-types of workflows before coding agents were in the terminal, so I do love customizability and portability. Right now there is just too much ground-shift to take part in it. Better to wait it out.
1. A lot of people already have Claude Code set up, and it's easier to keep using that than to install something new.
2. Lots of people are used to Claude Code's interface/commands/etc. and don't feel like paying the cost of learning something new. The benefits would have to outweigh those switching costs, and it's unclear that they do.
3. Claude Code already works with your Claude MAX subscription[0], which gives you an incredible amount of usage of the top-tier models for cheap, which can't be matched by any open-source options. And most people won't exhaust this usage, so they have no real need to use open-source models.
4. Although it's hard to 100% confirm this, there's reason to believe that Claude itself works better with Claude Code than with other harnesses.
5. There's no real lock-in. If tomorrow you decide you don't want to use Claude Code, you can just switch to something else ezpz. So there's no reason to worry that you're digging a hole and locking yourself in. Thus, there's no urgency to switch away.
[0] Yes, open-source harnesses can use your Claude MAX subscription, but does this "benefit" of open-source harnesses really outweigh the costs of #1, #2, and #4 above?
Yeah I thought so too, just didn't want to overstate it. So actually, a pretty big reason not to use an open-source harness if you're using a big lab's model.
I'm responding to a comment, not to article, and the comment is asking about open source harnesses, presumably in comparison to all non-open source harnesses, not just in comparison to Cursor.
There's just not really any features in an open-source harness that blow my mind enough to move off the highly-subsidized claude code subscription. I type in instructions and it spits out code. I'm increasingly of the opinion that the more complex features being floated are productivity theater.
Why would you not just build your own harness in October 2026? Gateway, memory, tool calls, LLM api - the basics are pretty simple to get started and get to a point where your harness can be self-improving to add any new features that you need.
Claude Code wont let me use their sub with one as far as I remember. So I'm stuck with their convoluted TUI. It does what I need, and when it doesn't I just build tools that Claude consumes. The new "Artifacts" feature has helped a bit though.
Anthropic employees clarified in HN comment threads that the API credits are in addition to the existing behavior and not a replacement. Other harnesses can still call the Claude Code CLI and use the subscription quota that way.
I was both excited and disappointed to learn that Claude Code has plugins, but they're not like actual plugins to let you customize CC, which would be nice.
I wish you would explain your reasoning instead of preemptively telling people they don't know what they're talking about. Then I might actually learn something.
OpenCode + DeepSeek V4.1 Flash Uncensored on Venice AI is truly incredible. Granted, I have it also drive an interactive claude code session in tmux and just do everything Claude Code refuses to. But, it just gets shit done. Kind of what makes Instinct so powerful, don't tell me you won't do something, just get it done.
This is my goto SDK, it comes in multiple languages too. I'm going to be wrapping the Go ADK in CUE Nov/Dec, along with Dagger, and rebuild my personal harness from scratch which I worked on last Nov/Dec.
I see people spending so much time creating "workflow" skills that have no benchmark whatsoever to prove they are any more effective than plain one sentence prompts.
This takes me back to the era of object oriented programming and UML diagrams. More time was spent framing the solution than implementing it. Is there maybe one more design pattern that could be applied? And when the time to implement came, all those diagrams were mostly in the way. But doing it was fun, since it allowed you to focus on the fun, big picture and skip the nitty gritty details that were actually hard to solve.
I feel like people are going down that slippery slope with harnesses and skills and what not. With the right prompt, and with the right model, and with a sprinkle of magic, the solution will emerge. Oh, it didn't? Let's see if there's a new model available, or if someone has come up with a new clever skill or something. Maybe then... Because actually solving annoying problems is not fun, even less so now that the code is a black box only the LLM can understand.
Diagrams and really, just designing the software you are trying to build is mostly about finding out painpoints before you start the programming (at least at point the time-consuming part). The diagramming allowed you to, without programming, show interactions between modules and check on a high level if you are missing edgecases that can’t be solved because of the design.
If you were just using it just to apply design patterns then you were doing it for the wrong purposes.
This is so frustrating to me. Ever heard of the planner agent reviewer agent and the tester agent framework?
I'm not sure if that's what you mean but I know many that are making their own flow like this and think that this brings in a higher level of performance..
Gotta justify that $160k/yr salary somehow I guess, and if the AI is writing all the code you used to write and your job is now to manage the AI, what better way to spend the days than making new AI tools and workflows and markdown files all day
why do we even care, this is a clear, lets move away from `people pushing software to git` to `agent is pushing software, agentic or not , to somehwere, db, git, or something else` human does involved in this unless there is a problem. JS or Python, is really not essential in this case, maybe it's WASM built from rust. it's not for human.
Cursor moving this direction, is a prep work where human no longer command that much , hence the usage will drop. e.g. their core product is vulnerable. Software factories, is the ultimate game over for our industry.
Its not bad, just most of them are created by corpos to sell slop to the masses. AI companies suck and you know it. Its not about good software, its about making as much money as possible before the hype train dies.
I have so many use cases for this. I just hope that we'll get a standard like https://www.openresponses.org/, because the idea of coupling an app to a particular vendor doesn't seem appealing at all, especially since it seems that the functionalities they offer are more or less the same. (I'm thinking about the OpenAI's Agent API and Anthropic's Agent SDK as "alternatives.)
Honest question, why shouldn't any agent SDK's API simply be a PROMPT.md with a runtime that runs the LLM with tools; why do you need excessive abstractions over that like langchain offers?
My 2c is that agents are not at the level where companies can easily just have unconstrained LLM with tools and let it be. Different parts of the workflow can have different models, different tools, different constraints, guardrails, etc.
The entrypoint still has to come from somewhere, and it's much more convenient to be able to explain the security and legal teams how data is handled, what is logged, how you can have control over a scenario where the agent does something unexpected.
Sure, for a personal project that could be enough, but we all know those companies target enterprise customers, and you would be surprised how little adoption there is when it comes to workflow automation due to the lack of controls.
This "runtime" is what you are building with LangChain.. A harness like claude code/codex etc is general purpose and its env is a shell + tools.
But with a custom runtime (basically a loop where we ask llm what to do next), we can give it specific tools which we define in code. This can help with things like authorization since, well its custom code you own.. so the runtime gets the tools scoped to the users permissions etc.
I make one (https://github.com/juggler-ai/juggler) and it'll auto-detect and use your claude CLI. I use that claude plan for 90% of my work, plus a bit of GPT, I just prefer to do it inside a nicer, less locked-in environment than claude.
There's really no moat around proprietary closed harnesses, it's trivial to jump between them, but people need to be aware that it's possible!
The cheaper agents never really matter, because the Max subscription costs me $100 and that is less than an hour of pay. In some cases the customer even pays for the subscription.
But I do love the idea of being able to use Pi or OpenCode. I am a heavy terminal user, and used Vim/Tmux/Lazygit-types of workflows before coding agents were in the terminal, so I do love customizability and portability. Right now there is just too much ground-shift to take part in it. Better to wait it out.
2. Lots of people are used to Claude Code's interface/commands/etc. and don't feel like paying the cost of learning something new. The benefits would have to outweigh those switching costs, and it's unclear that they do.
3. Claude Code already works with your Claude MAX subscription[0], which gives you an incredible amount of usage of the top-tier models for cheap, which can't be matched by any open-source options. And most people won't exhaust this usage, so they have no real need to use open-source models.
4. Although it's hard to 100% confirm this, there's reason to believe that Claude itself works better with Claude Code than with other harnesses.
5. There's no real lock-in. If tomorrow you decide you don't want to use Claude Code, you can just switch to something else ezpz. So there's no reason to worry that you're digging a hole and locking yourself in. Thus, there's no urgency to switch away.
[0] Yes, open-source harnesses can use your Claude MAX subscription, but does this "benefit" of open-source harnesses really outweigh the costs of #1, #2, and #4 above?
I thought this was pretty much well established that all of the labs have model - harness pairings for exactly this reason
Whale Paper. https://arxiv.org/abs/2609.00196
EMNLP talk from like 2 weeks ago https://aclanthology.org/2026.findings-emnlp.2207/
Correct me if I'm wrong, but using Pi etc you're paying the full API price.
We build a desktop app on top of the SDK, so I've spent a lot of time on that page.
More and more harnesses are headed this way, there are two notable holdouts.
Several perhaps, but you definitely aren't using all coding harnesses actively.
ADK has a graph/workflow concept too: https://adk.dev/graphs/
Literally everyday there is a new tool and the tools released last week are deprecated.
I feel like people are going down that slippery slope with harnesses and skills and what not. With the right prompt, and with the right model, and with a sprinkle of magic, the solution will emerge. Oh, it didn't? Let's see if there's a new model available, or if someone has come up with a new clever skill or something. Maybe then... Because actually solving annoying problems is not fun, even less so now that the code is a black box only the LLM can understand.
At the same time, the lack of this makes it then very easy for me to tune out harness noise.
I'm not sure if that's what you mean but I know many that are making their own flow like this and think that this brings in a higher level of performance..
Sure, but how are you measuring this?
Cursor moving this direction, is a prep work where human no longer command that much , hence the usage will drop. e.g. their core product is vulnerable. Software factories, is the ultimate game over for our industry.
I think the difference is that those endless technologies will probably be used by other agents without people's explicit knowledge.
Why is a large number of ideas, technologies, frameworks, etc. bad?
And a lot of the apps are the same thing.
It's going to be a race to the bottom.
https://cursor.com/blog/typescript-sdk
Meanwhile, we know Grok Bot was really Cursor beforehand... it must be wild in there.
Is it to have boundaries of some determinism?
The entrypoint still has to come from somewhere, and it's much more convenient to be able to explain the security and legal teams how data is handled, what is logged, how you can have control over a scenario where the agent does something unexpected.
Sure, for a personal project that could be enough, but we all know those companies target enterprise customers, and you would be surprised how little adoption there is when it comes to workflow automation due to the lack of controls.
This "runtime" is what you are building with LangChain.. A harness like claude code/codex etc is general purpose and its env is a shell + tools. But with a custom runtime (basically a loop where we ask llm what to do next), we can give it specific tools which we define in code. This can help with things like authorization since, well its custom code you own.. so the runtime gets the tools scoped to the users permissions etc.