I don't really get this. At this point, my limiting factor is not how quickly Claude can self-trudge through code. It's whether Claude is going to do the task correctly or not.
I need more mechanisms for controlling long-running sessions and dynamically injecting my thoughts, correction, and nudges rather than faster ways to burn through my tokens without knowing if the results are going to be correct.
"Agents address the problem from independent angles, other agents try to refute what they found, and the run keeps iterating until the answers converge."
So you will be supplying the "ground truth" (test suite, detailed spec, whatever) and empower an agent to use it to guide the other agents. Currently a lot of people do this sequentially in the form of multiple code-review passes by fresh agent sessions looking at the work of previous sessions.
Adversarial models are a longstanding technique in ML so it makes sense they would try to go this way.
This reminds me of something I do manually all the time (tell Claude to do something, ask another instance what the first one did wrong, etc), and it has given me good results. If the harness can now automate that, it seems great!
I don't know, maybe I'm doing it wrong but I feel LLMs add a slop debt, and each agent pass just exuberates it.
Like I had an LLM implement a spec and said it was done... Except it had a ton of `casts` everywhere. Okay, my bad, I should have been clear "NO CASTS", so I use the LLM to remove the casts, except it just kept making things more and more complicated and ugly.
It took me taking a break and having a shower thought to realize all the ugliness is because one type should have been broken up into 2, which would remove a ton of generics and code. But Claude never suggested that, it was always "we need at least one cast here, or we need 1000 LOC of generic factories". I tried multiple new sessions with various prompts too.
Maybe one day soon LLMs could pay off their own slop debt but at least right now I don't trust them to write code unseen.
Edit: Maybe the correct action should have been to delete everything and make it re-write everything from scratch with the clear "NO CASTS EVER" rule. But still the point is feels like having LLM clean up after an LLM doesn't work well enough to just have keep it in a loop and never look at what it does.
I've had to put a fair chunk of effort in to skills that will run deterministic mechanisms to unslop a codebase (cyclomatic complexity grading has been really helpful here) as invariably some amount of guidance around principles will be missed over time. I've found it does help, though. Certainly I'm getting overall better results from Flash and Sonnet over multiple runs for fairly modest token increases. GPT 5.5 less so, but that's because it scores better in a first pass. I won't really know until I gauge it at the end of my sub month which has been more cost efficient for me all things considered.
Write a skill outlining your expectations of the code, put that skill into the pipeline, so that it can be included within your workflow.
Webdev here, but currently I have:
- a skill where I outlined how the architecture of the system should look like, with guards (static analysis, architecture tests, linting) confirming that the code it generates adheres to standards
- a skill that tells it how tests should look like (use generators, write both feature / unit tests)
- a skill that tells it to generate docs from the code in a form of acceptance criteria (Given / When / Then)
- a skill that tells it to generate frontend uat tests + accompanying backend seeders given the AC
- a skill that tells it to verify that ticket objectives match what was delivered
At this point I still need to guide it to move task from one stage to the other (coding, testing, verification that indeed what was coded adheres to what was required), but I believe that these dynamic workflows can automate this work as well.
I’m in a similar boat. I find that longer sessions will introduce “noise”. I have to be extremely explicit to avoid adding this noise, as it pollutes the future output of the models. Sometimes it’s innocuous, other times it can derail sessions as the 2nd or 3rd pass introduce even more of their own noise.
To me, it seems the models are inherently designed to do this. Creating more verbose output than input, generating plans introduce things I didn’t ask for, extras, more “defensive” code that makes sense at first but is completely unnecessary in practice… I find it exhausting, but it’s important to pare down the output / plans at each stage and trim the generated stuff that isn’t needed.
The problem is that we have an ever growing and large number of constraints, and not following even a single one means the result is sloppy.
I don’t see them fixing this any time soon, and thus human in the loop is a requirement to use these tools effectively. That is unless you love your slot machine dopamine rush enough to ignore quality gates and respect for your peers time.
How would you prevent a junior engineer doing this mistake? Presumably, you would setup a lint rule. Do the same for LLMs. Run the linter after each edit through a hook, give feedback to the LLM. Write your lint rules with clear explanations of why the behavior is a problem, and nudges to the good behavior.
You wouldn't prevent the junior from making this mistake.
You would correct them once or twice, and they won't make the mistake again.
It's something we can't do with LLM's currently, so we all just try(and fail) to predict any possible failure ever, and then somehow try to cram it into the limited context.
I've been reading writing Rust for a long while now, since before 1.0. I'm capable of critically evaluating Rust code. I'm also a happy Claude Code user, mostly for lightweight uses like generating scaffolding, prototyping, and debugging.
The pure LLM, no human intervention vibe-coded PRs on Bun since the vibe-rewrite to Rust contain the worst coding horrors I've seen in 20 years of programming.
Setting aside the quality of the change itself (I would have done it differently, for sure: it is pretty straightforward to build a safe abstraction out of this type), the utterly pointless "source-text consistency test" added here is easily the worst example of "test repeats implementation" I have seen in my career:
> Currently a lot of people do this sequentially in the form of multiple code-review passes by fresh agent sessions looking at the work of previous sessions.
Up until now I've used a review loop approach, where within a Claude Code session I just tell it to spawn three review sub-agents, each with context of what's going on and instructions to look over all of the changed code in search for serious/critical issues, but otherwise a more fresh look at things. It works really well for the most part (token usage aside): https://news.ycombinator.com/item?id=48277011
Ground truth is not consensus, it has to be graded against what actually works for the original goal. Plenty of scenarios with AI and Humans can result in consensus around incorrectness.
While pedantically correct, I think the comment above assumed that you've correctly specified the work. If you can't correctly specify your work, AI agents are just going to help you get a non-solution faster.
Isn't coding the act of specificying the work to a processor? And AI agents are supposed to bridge the gap with intelligence from less specificed to more specified or possibly even more intelligent and alternate implementations?
What I meant by "ground truth" is that it is not fuzzy, not AI-evaluated, and not a consensus. The test suite passes or it doesn't. The codebase lints or it doesn't. The performance improved or it didn't.
An agent can help you create the specification, but it's up to you to know whether it's correctly testing that you got the result you wanted.
Yes, that, accuracy, speed, and single computer-use.
I find those to be the limiting factors to speed.
I have extensive rules, I do extensive planning. Yet at implementation, the rules are not respected, errors are introduced, etc...
I spend more time fixing than writing code.
Then speed... Because of the fixes and bad code quality even with frontiers model speed makes a very big difference. I (agents) spend hours daily doing reviews and fixes. 5x speed boost would make me much more productive.
And when working super fast with agents, having only one computer is limiting. Even worktrees don't solve problems because I use things like convex, chrome use, etc... and it conflicts with each others all the time.
Still many problems to solve. It's already evolved so much in the last two years.
This is my experience. Quantity of output is not the issue right now. Quality is. But I’m not sure if this will ever be solved for, given LLMs are non-deterministic sophisticated autocomplete at their core.
Sure, ‘human in the loop’ and all that jazz, but I feel like my knowledge suffers even with this approach. I have to use llms w pinpoint focus to get decent results.
The original copilot completions behavior might be peak llm performance for coding, sans having an agent write boilerplate and such.
A more interactive Claude code would be great instead of 50 “here’s a tiny snapshot of a change shorn of the context you need to understand it. Yes or no?”s
When this is all finished and done, these coding models will allow you to rewrite the linux kernel in rust, recode Kubernetes in assembly, and create your own web framework in 10 min.
But each prompt will cost your company, 10 to 15 million dollars. An extra 20 million if you ask them to review the code and improve the comments.
I think for now it's better to convert tokens into code/library code and then work with that for deterministic results rather than relying on Claude being correct or not.
yes I agree with this, more granular going back, letting me interrupt where it went off the rails, or even editing file reads myself etc would be lovely. Ingesting parts of other conversations would also be cool!
Not with those exact terms, but it is certainly being discussed. Wes McKinney said in a recent talk that with current coding agents there’s no longer an excuse for shipping suboptimal code that takes on tech debt. Writing tests has never been cheaper, writing custom fuzzers, linters, and other harnesses that serve as guardrails has never been cheaper. His take is that “we didn’t have enough engineering time to do it right” is no longer an excuse, and the only excuses left are that you don’t know any better or you have bad taste.
Dynamic workflows, in my experience, make Claude more effective at complex long-running tasks. They help precisely with getting Claude to do the task correctly.
It feels more like a bespoke build system for the specific task/project than prompting a freeform chat.
As long as agents are fuzzy (which they will continue to be with the Transformers architecture), the need to validate will continue to exist. I cannot imagine merging code without at least 1 human review.
Interesting to note, not sure if this was known publicly before today's blog post:
Rewriting Bun with dynamic workflows
An example of what dynamic workflows can unlock at scale is the recent rewrite of Bun. Jarred Sumner used dynamic workflows to port Bun from Zig to Rust with 99.8% of the existing test suite passing, roughly 750,000 lines of Rust, and eleven days from first commit to merge. One workflow mapped the right Rust lifetime for every struct field in the Zig codebase. The next wrote every .rs file as a behavior-identical port of its .zig counterpart, hundreds of agents working in parallel with two reviewers on each file. A fix loop then drove the build and test suite until both ran clean. After the port landed, an overnight workflow addressed unnecessary data copies and opened a PR for each for final review. While not yet in production, all of this was handled by dynamic workflows. Jarred will be writing about this more in the future.
I'm extremely skeptical that dynamic workflows had anything to do with this. I've been able to refactor one of the most complicated parts of our code base with similar results.
Mechanical refactors are relatively straight forward for agents.
> I've been able to refactor one of the most complicated parts of our code base with similar results. Mechanical refactors are relatively straight forward for agents.
A rewrite of bun in Rust is unlikely to be a trivial mechanical refactor. And if you are not sharing what the complicated parts were, or how big it is, how do we assess that the task was similar?
Unless you are intimately familiar with the bun codebase and you've already made that assessment.
I am getting so confused when to use what... agents, sub-agents, tasks, team mates, /goal, /loop, and now workflow. Each with different degrees of effort.
Don't make me think.
All these knobs are also exposed in ChatGPT, which I am more familiar when chatting. Which one of the models? Do I go Instant, Thinking, Pro? Extended Pro? Oh no, maybe I need Deep Research.
Sometimes I think it's on purpose. I fear if I try a lowest knob, it will miss something. So turn everything up. And token usage goes up.
i think it is on purpose, but not for the cynical reason of burning more tokens. developers like knobs, it helps us feel like we're in control. even when we're not.
so the ai companies give us knobs and buttons and sliders to make us more comfortable.
The #1 goal for Anthropic and others is to take the longest running process possible and make it entirely opaque to the developer. It's the only way they can build a moat for a commodity. I would highly recommend building your own multi-stage orchestration flows because then you'll get a much better idea of where you need to be in the loop, and where you can save money. Once entire organisations are functioning only as extensions of Anthropic, they'll put the prices up and squeeze the shit out of the market.
It feels like we're far past the point of where having AI do more faster is helpful.
It's telling that they used "rewrite Bun in Rust" as the proof point here. It's cool! But the vast majority of software engineering doesn't start with tens of thousands of tests, where making them pass is the whole job.
In my experience, AI still drifts from what I meant it to do on anything bigger than building a widget. My time is spent suspiciously reviewing output for changes the agent snuck in, or invariants it broke. I talked with a friend recently where the agent broke the test harness badly enough that none of the tests mattered for 3 weeks. They did pass, though, so CI never complained.
There's something at the intersection of context engineering, managing that sloppy pile of markdown plans, and good old fashioning system understanding that's the real bottleneck.
"In my experience, AI still drifts from what I meant it to do on anything bigger than building a widget."
I've had code bases with tens of thousands of lines of code built from scratch that I hand-reviewed every line of and worked with the AI to improve, and haven't had this issue. I feel like a significant part of this is due to an involved /plan stage -- going back and forth on building out a plan for what you want the AI to do involves surfacing the assumptions that you would have called drift if you asked them to implement it directly from your prompt.
Once the plan has been refined and is what I want it to be, getting it to implement everything in TDD style has for the most part given me 100% working code, as I wanted it to be, without issues. It definitely helps that I'm a principal-level engineer with extensive architectural experience -- but if you're able to tell the AI in detail what you want, have it ask questions for clarifications, and read through a plan before getting it implemented, and have a solid testing plus manual qa process (automated by chrome devtools mcp) in place, I've find that you can one-shot complex features, rewrites, and even not-insignificant applications that would have taken days to write by hand in a few hours.
Depends - using Sonnet here and generally it should be as you say: plan would produce the result.
Still Claude will sneak things in - in my recent plan, for example I had defined, per acceptance criteria what colours the statuses should be: green for live, blue for sold, grey for anything else; it changed this to: green for live, orange for in progress, blue for sold, red in demolition, etc. When pressed why did it to this, it was unable to explain why. This is with a plan where AC were explicitly provided from the task in Given/When/Then format and were to be adhered to strictly. I've caught this within planning, but I shouldn't need to be doing this.
Even in standard prompts where I tell it "Change this label from X to Y", it ended reordering the tabs unrelated to ask. Again I was not able for it to explain why - it was so abrupt. And it was in fresh context, without any pollution on what I expect it to do.
I also noticed a different behaviour regarding skill; today and yesterday it would not be following skill guidance at all ie: skill writing skill - I'd have to explicitly tell it to test skills after writing them, when this is a behaviour expected by default. Similarly with other skills - knowing that it should have done something per skill guidelines and it not doing it at all. This is new behaviour that I've not seen a week ago.
There are certainly domains where AI is not so effective, but at this point I would agree that at least in terms of web development if you can't get effective results from agents at this point it is a skill issue. That skill can be learned, if you recognize that learning is part of the solution. I do think prior experience in product design, specifications & business analysis as well as engineering leadership are all extremely helpful. Its about putting the agent in a box so small that it really can't screw up; but its also about being able to review design and code rigorously - to see around corners and anticipate possible weaknesses etc. There is really nothing I have to do when working with an agent that I haven't already been doing for decades but it seems to me that a lot of developers have never found a single bug while reviewing someone else's code.
A few of us from the Claude Code team will be hanging around if anyone has questions! Very excited for this launch -- dynamic workflows have been a game changer for engineering here at Anthropic. Can't wait to hear what you think.
Using the keyword “Workflow”like “Ultrathink” is problematic?
Ultrathink is uncommon enough that it is unlikely to be used in code or prompt outside its intended purpose.
Workflow is generic keyword and used in so many contexts both inside the codebase and orchestration tooling like say temporal.io or others that name their constructs “workflows”.
Hi Boris. Love the velocity of features. Are you planning on adding a secrets manager? Enterprise workflows almost always require an encrypted parameter or calling a secret.
Thanks to you and the anthropic team for developing such exciting tools! The blog post seems to position workflows for “breadth”: generating fixes / refactors against large code bases. What about for “depth”: developing specific new features and functionality end-to-end? I’ve struggled to make this work reliably using the current experimental agent teams. Does this replace or augment that functionality?
Yes, it also helps! That's a place where raw model capability is the most helpful, but we do find that some dynamic workflow configurations can be helpful too.
Cool! If you can point to any examples of those types of workflow configurations I’d be super interested. For example, to have a team of agents review a PR and iterate on it until all requirements are met including UX, security and product functionality goals. If they could “converge” to a solution like workflows seems to be designed for that would be amazing.
How granular is the control over the internal process?
In my experiments I've had some success modeling the work to be done as a DAG of typed artifacts with a combination of code + LLM doing decomposition, transforms, synthesis, and fitness checking to generate the output. It took me a lot of tries to arrive at that formula and it would be cool to have something more general. I also run part of it against local compute because it would be far beyond my budget to do it all on Opus, so something for that would be nice too.
Is there an example of how y'all use Dynamic Workflows internally that you could share with the rest of us here so that we can mimic something similar?
Hey, yep. A few things I personally used dynamic workflows for over the last few weeks:
1. Autonomously landed 20+ optimizations to reduce Claude Code's token usage by ~15%
2. Ported tree-sitter, color-diff, yoga-layout, and a number of other WASM and Rust native modules to TypeScript, improving CPU and memory use by 2-10x in the process
3. Made our CI faster, and repeatedly found and fixed flaky tests (with /loop)
4. Migrated from regex-based bash static analysis to tree-sitter, reducing false positive permission prompts by 45%
5. Reduced Claude Agent SDK startup time by 61%, by repeatedly profiling and optimizing the startup path, putting up a number of PRs in the process
> Ported tree-sitter, color-diff, yoga-layout, and a number of other WASM and Rust native modules to TypeScript, improving CPU and memory use by 2-10x in the process
Curious to learn more on this (unless there’s a write-up in the works). I’m naive on this matter but:
1. is this because it’s higher cost when passing objects back and forth across the JS/native boundary?
2. Does this have anything more specific to do with use of Bun?
3. is the stance for claude code then to keep all the deps in raw TypeScript?
4. How do you folks keep these ported deps up-to-date?
Very cool. What % of the CC team's engineering would you say goes into QoL (as opposed to new feature development)? Obviously some live in a grey area, while others are more clear like making CI faster.
Is there not a reason to instead port claude code to rust? Do you have internal benchmarks that show that claude code is better at typescript than rust?
just wanted to say thank you, just did a 2 days "ai computer use" workshop - think a virtual desktop on hetzner with claude code in yolo mode, a github account, vercel and logged in into a google account and claude had all the credentials and then let a mix of marketing / product manager / sales / customer support let loose. 2k token budget ... and just let them see do magic again and again.
VSCode has an official client? Given IDE usage is being restricted from Claude Code via the CC SDK tokens going to the Claude API rather than your CC Subscription, i'm unclear which IDEs can actually use claude code now.
Eg is Zed capable of using a Claude Code Subscription?
Oh, yea here's all the proof you need. Even Zed themselves admit you won't be able to use Claude Code via ACP via Subscription: https://zed.dev/blog/terminal-threads
So yea, bcherny didn't reply to me but as far as i can tell - No, Zed nor VSCode will have Claude Code natively in it. The best we can do is embed a Terminal into the editor and run CC in that.
With that said, because bcherny advertised VSCode, i'm going to guess VSCode is going to get special treatment. Really annoying.
to be clear, i'm referring to the recent fact where it appears that they're disabling all Claude Code (Subscription) usage from the SDK. Which ACP would be included on.
As usual though it's not super clear exactly what is allowed or not.
Can you please fix the issue where like 99.99999% of the time Claude tries to launch a subagent on its own accord it gets "Prompt is too long" and tries several more times, then gives up and does it without the subagent. Big waste of time and tokens and not getting almost any subagent advantages. Not kidding that this happens about 100 times a day.
I tried creating a workflow in Claude 1.9255.2 (1dc8f7) 2026-05-27T01:57:20.000Z
and got
API Error: 400 messages.3.content.11: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.
Tried again in
Claude 1.9659.1 (193bcb) 2026-05-28T16:22:15.000Z also but may need a new chat
This is really dissapointing release for such a promising technique. Long walks with fanned vectors can actually be token optimizing vs token burning when combined with self grading each agent along the walk and compared to manual long coding walks to solve first pass problems. But instead this frames it (assumptively) as a tokenmaxxing strategy. There are also many other strartegies that can prove effeciency and wider solution consideration with consensus, but none of this is explained why its an improvement or better than other technqiues.
Its like you guys aren't even aware of the primary problem you are all facing: your token burns aren't paying off anyore against standard coding -- and looking net negative. I have to ask, are you this unaware of your core problem set here?
There are no any examples, proofs, or scenarios that show why there is improvement either in complexity or reliability of the solution or effeciency to the path of the solution. I'm baffled.
Quite a thing to use Bun rewrite to Rust as example of dynamic workflows, while now it is considered as anti pattern which leads team to stop supporting the tool due to inability to properly understand and navigate 1m vibe coded Rust lines
I'm going to be honest, this very much reads like an exciting new way to burn up as many tokens as possible. Large amounts of parallel agents that all have all their work double-checked by multiple other agents, and that keeps running for a longer period of time?
I feel like there are more efficient ways to tackle the issues given.
Possibly. But otoh one cannot complain that agents don't produce high quality code while at the same time not allowing them to thoroughly go through all the steps required to produce high quality code
I tried this out yesterday - lucky enough to have access through EAP at work. The workflows that are generated are quite good - smart parallelisation and phasing. End results for larger chunks of work are also much better, which I attribute to more of the work having clean context windows (Opus 4.7 is unusable past 200k conversation length, and each subagent ends up using less than that IME). They also seem to have a validation phase hint in the workflow generator which also helps a lot. Speed is a bonus.
You can achieve a similar result manually prompting to use subagents, yes. But the TUI for in flight dynamic workflows is really nice - great visibility into exactly what's happening.
Honesty, for anything larger than a 1 shot PR, it's worth firing off a workflow for better automatic context management alone (more work done in the first 20% sweet spot)
Are these “features” just hooks to get people to burn more tokens faster?
I’m at the point where deciding what we should and should not do takes a lot more time than actually doing it. More agents just means running faster in potentially the wrong direction
They’re pure enterprise features - needed for massive legacy codebases with tens or hundreds of similar enough coding tasks - where there is a lived “cost” of not doing this type of work paid by every engineer working around it
Not sure why Claude does not have AskUserQuestion implementation that works for spawned sessions: subagents, teams, workflows. Without it, spawning hundreds of subagents and wait for final result without single input feels a bit risky.
Here is the solution to it. Built on a SQLite DB and MCP, blocking until the question is answered, supporting all possible question types, with a CLI or web interface for answers, `ask_human_question` fills the gap in efficient subagent management.
wrote something similar for my own use/work stuff; seems everything is converging towards similar ideas.
IMO, this style of workflow/agentics is how all SWE'll look like long term. Automate everything into a big pipe-y thing. How it's gonna be modelled is up in the air though. lots of different approaches:
I made my own knockoff of that for myself https://github.com/vblanco20-1/AgentLoom (not really usable, just a vibecoded prototype), based on the workflow files found in the Bun repo. Ive been using it but pointed at deepseek flash to do some really large scale stuff. Its a fun way of using agents, and highly useful for tasks like code review to apply some rules, or to find vulnerability candidates. Funny enough, i used it in the same way claude does, vibecoding the workflow scripts and prompts themselves.
I did find it uses tokens like crazy, i migrated Pixel Dungeon (java) to C# as a experiment, and it used almost 2 billion tokens. It was just 20 bucks due to deepseek flash, but i shudder thinking of how much money this uses when run on the real claude API pricing.
just to test the tech. No real usage other than for the fun of it.
I did port stb_image from C to Jai which i was able to fully verify and harden and that one ill give more use. Im also using the same workflow system to perform agentic translation of a game i work with from english to various other languages, the results are far better than the commercial "human" translation services we tested. And i also use it to fix OCR issues on PDF books im ocr-ing for a data pipeline. This kind of workflow/wide agent swarm system is rather useful for many things where you want to "apply" the same prompts across a whole codebase or just in parallel.
Anthropic is going to price themselves out of code, but still find a nice market providing service to senior management. Their long term play is virtual employees rather than tools for humans.
So far Codex /goal has been amazing but Claude Code /goal or even /loop does not work hard enough and gives up. I have observed it just claiming it’s “iterating” in a broken loop or simply giving up.
It's the later. You can view it and see fine grained progress, but you can't interact with it. I hope that's coming next, because it would be useful to steer later phases or even agents
Absolutely annoying to have it assuming that I want to use this when I type workflow in the prompt. Like thats not already a thing in half of the software projects
“We realized the tech is not as addictive as we’ve hoped so we won’t be able to raise token prices enough to be profitable, so here’s a way to make you consume a lot more tokens without even realizing”
I need more mechanisms for controlling long-running sessions and dynamically injecting my thoughts, correction, and nudges rather than faster ways to burn through my tokens without knowing if the results are going to be correct.
reply