> Your Claude subscription—which is a cheaper version of the Anthropic API—is restricted to use with the Claude Code CLI/Desktop, Claude CoWork, or @Claude in Slack.
Thats not true at all. You can use the Agent SDK [1], which uses your subscription [2]. I use it via ACP [3] with custom system prompts and tooling. I have found it very powerful and flexible. It has its own agent loop, of course, so maybe thats the limitation using it with opencode?
[2] They were talking about giving credits for the SDK usage rather than it using your allowance directly, but that seems to have been put on hold for now. If and when that changes, I will likely jump ship, but I am more than happy with it right now.
[3] there isn't an official ACP wrapper - zed have one but its quite limited. its trivial to build one though, or you can just use the SDK directly and wire it into your interface of choice.
The fun part is they planned for this to change to API pricing. When they announced that I started shifting off claude code immediately so I could keep my custom tooling. Damage is done. I won't build up around claude code or agent sdk any longer because I expect this to change. The announcement says they're "pausing" the change so expect this and avoid the lock-in!
A lot of those tokens are traversal - either searching for code or following call-sites. Basically building enough context to be able to work on the task.
You can reduce a lot of the token use for traversal by giving your agent access to some form of LSP in addition to hierarchical direction with your AGENTS.md (or equivalent) for monorepos - but a spread-out codebase is always going to end up requiring some form of traversal to solve each task.
And that traversal isn't just token use - its repeated round trip latency (LLM (queue time -> prefill -> decode -> output) -> Agent (parsing -> tool call -> tool response) -> back to LLM) for EACH step (well, some can be done in parallel, but in practice its mostly sequential) - slowing down the task considerably.
Locality and structure are key when it comes to efficient use of agents. The context window is always bounded and attention across it is inconsistent.
That’s great! But don’t fall into the trap that the graph is the ultimate representation of all. You will get a topology, but you often need more than topologies to understand the whole. Behaviour of the parts themselves is covered in the manifesto, and there is a companion that covers the connecting of the parts: https://shapeofthesystem.com/the-shape-of-the-whole
Yeah, you need actual models to understand what is going on.
Basically, when I was using my first version of Coherence to write down what system does on lowest level I found myself imagining in my head database records. And this is exhausting when you have more than 2 tables to think about.
And so I don't want to imagine them, I want DB records represented as entities I can simply render onto screen and look at them while I think how they should interact with each other.
Now, this is topology + domain modeling. Topology/semantics here is describing what happens with those domain models over time, what states they, how they transition, what attributes models might have, and most important — which code symbol describes all of that.
So for example, I am making Job Tracking System. I need to persist Job record. I visualize it as a row in table and I can model it without attaching to code. And then we can say:
- "this entity (job) --(has_attribute)--> status", or
- "job -(described_by)-> class JobRecord",
- "job -(observed_in)-> trace",
- "job -(verified_by)-> job_test.rb".
This gives us useful semantics / description / context / domain knowledge + links it to code + test that confirms model shape, or system state transition, or process, or product behavior.
And you can describe it at any level you want.
If you don't know the implementation details yet, a greenfield project? Well, just start with Product level specs, high level description what users will see.
Is it brown-field legacy monolith 500k LoC? Well, just reverse-engineer intent from code (I even have guide how to do this and I did it on 10k LoC successfully), then align people on that "intent graph", and then connect it into CI/CD so that it becomes default review artifact on EVERY single pull request.
And the key here is to isolate context around domain model we're interested in. Make it minimal but useful to make decisions. And make it easy to pull more [relevant] context, or remove [unnecessary] context.
I still don’t fully know how to avoid tunnel vision once you have only one slice of the graph and models in front of you. E.g. failure mode might be somewhere not even described by a link or something exists entirely outside of the graph.
This would be the next thing to solve.
And meanwhile I'd like to thank you for showing me this problem.
Thanks for the honest feedback. I genuinely wish I was a better writer. The posts in particular are all quite formulaic, but the idea is that they just provide a narrative access to the manifesto itself which is where the real meat of the argument is.
I would please urge you to read further into the manifesto itself but would also recommend you start at the foreword so you can understand the reason for the use of AI assistance in my writing.
Alleged author here - on my regular account. Thanks for the feedback. I have been quite upfront about the use of AI in the foreword. I have severe inattentive ADHD and have used AI to take my writings and present them in a way in which I feel are much more coherent overall.
The actual criticisms you have about the content however, I'd like to challenge:
The "adding up to one" is just a simplified gloss over softmax. It's very possible it reads poorly, and thats on me - not LLM gibberish.
As for the incoherence - I have to totally disagree. You have merged the 2 things the post keeps apart - capacity and attention over it. That a model can swallow a schema and write code is a competence humans share. We have been doing it for decades. Besides, the claim was never about us sharing capacity (other than stating it is always bounded) - it was about our attention failing in eerily similar ways.
So, AI slop, no. AI assisted, absolutely. It's sad that some judge the "who" more important than the "what" - especially for this kind of writing. But it's fair feedback nonetheless. I'll see what else I can do with assisting my delivery.
I think the problem with "AI assisted" is that it triggers the same alarm bells as AI slop, justified or not, so many readers just turn off. I personally rather read "shitty" writing than AI assisted writing.
Conveniently forgetting how they removed the jpeg-xl support from the chrome codebase despite overwhelming developer backlash that they then proceeded to ignore for over a year.
They literally tried to kill it - stating (nonsensical) reasons why it was obsolete and unneeded.
And since now the rest of the world have adopted it despite Google, they have crawled out of their slime pits praising themselves for its development with only a passing mention of cloudinary?
People bragging that they "dont touch code" and only "argue" with agents are reinventing the slowest possible IDE.
Obviously the agents are great at producing large chunks of code, but they often make minor and sometimes trivial mistakes which need amending.
Typing something like "in src/auth/session/token_manager.ts the refreshTokenExpiry variable should be refresh_token_expiry. update every reference and make sure nothing else changes" and waiting for the LLM to do its thing takes longer than opening the file and doing the rename yourself.
If you are describing microscopic edits in natural language you are not avoiding coding. You are coding through an extremely verbose, lossy interface with higher latency and lower precision.
Codex and CC are actually getting better at reviewing code and flagging issues. False positive rate dropped fairly significantly. Also obviously might be very personal preferency but creating clear specs and iterating on specs really helps to crystalize the approach I want to take to solving a given problem.
In the latest commit, refreshTokenExpiry should be snake case. Fix and make a note to do that in $LANG.
Or otherwise scope the rule. The point is you don't need to be super verbose about it and you get to fix forward too, preventing the same issue reoccurring. You build knowledge and context that let you move faster.
Thats not true at all. You can use the Agent SDK [1], which uses your subscription [2]. I use it via ACP [3] with custom system prompts and tooling. I have found it very powerful and flexible. It has its own agent loop, of course, so maybe thats the limitation using it with opencode?
[1] https://code.claude.com/docs/en/agent-sdk/overview
[2] They were talking about giving credits for the SDK usage rather than it using your allowance directly, but that seems to have been put on hold for now. If and when that changes, I will likely jump ship, but I am more than happy with it right now.
[3] there isn't an official ACP wrapper - zed have one but its quite limited. its trivial to build one though, or you can just use the SDK directly and wire it into your interface of choice.
reply