In this YouTube conversation, Robert Ta introduces Anthropic’s new "Dynamic Workflows" and Claude Code feature, where Claude can generate JavaScript harnesses on the fly to spawn and coordinate isolated sub‑agents (different Claude model variants) with their own contexts, goals, and token budgets. He explains that this solves the problem of "agentic laziness"—Claude prematurely declaring a task done—by splitting planning and execution across sub‑agents. Robert outlines six workflow patterns recommended by Anthropic: (1) classify‑and‑act, (2) fan‑out‑and‑synthesize, (3) adversarial verification, (4) generate‑and‑filter, (5) tournament‑style with judges, and (6) loop‑until‑done. He describes how to trigger workflows using the "ultra code" keyword, loop and goal commands, and plain‑English token budget specifications, and notes that workflows are saved in the .claud/workflows directory and can be shared via skills. A real‑world case study—Botten’s large‑scale Rust migration—illustrates cost and time implications. Jonathan then adds practical engineering advice: inventory existing repository tasks against the six patterns, configure the environment so tasks automatically trigger appropriate workflows, simplify skill definitions, and create domain‑specific rubrics from good examples to guide Claude’s output. The discussion blends technical detail with candid commentary on Claude’s behavior, offering a roadmap for developers to harness dynamic workflows for reliable, enterprise‑scale AI productivity.
Chapters
Anthropic published this article, A Harness for Every Task: Dynamic Workflows and Claude Code. The thesis verbatim is that Claude writes its harnesses on the fly, custom built for the task at hand. I kind of think of it like Tony Stark in Iron Man. Now instead of just Iron Man suit, it's like, hey, I'm just putting on new suits on the fly, on the go. Like you want me to be accountant? All right, accountant suit. All right, want me to be a product designer? Product designer suit. And that's really interesting because they're going for obviously enterprise scale and enterprise productivity with all of their plays going around right now to compete against OpenAI to win the enterprise race. So how this works is Claude writes a JavaScript file with special functions that spawn and coordinate subagents, isolated Claude instances with their own context windows and focused goals. The workflow decides which model each subagent uses, whether it be Sonnet or Opus, whatever, and it could run subagents in isolated work trees. The real problem this solves for AI builders is when you ask the default Claude code harness to do a task, it needs to both plan and execute In the same context window. Workflows split those across isolated subagents and combat agentic laziness, is what they say. When Claude stops before finishing a particularly complex multi-part task and declares the job done after partial progress. That's sort of funny that they call it agentic laziness instead of just calling it a lie, which is what it is. Like, it's like, Claude, if you're telling me you're done and then I look at your work, it's not fucking done, bro. It's not agentic laziness. Although we might call it that as well. You're fucking lying to me. Like, come on. This is really interesting. How do you actually drive it? Say, use a workflow or the trigger word ultra code to force one. And that combines with the loop command and the goal command for a hard completion target with a keep going sort of command. And you could also set token budgets in plain English, like only use 10,000 tokens, only use 500,000 tokens kind of thing. It saves workflows in the .claud/workflows. directory, or you could share them via skills. Some real use cases that have used this— Botten recently did a rewrite to Rust, and that was a large migration refactor. And I think that cost them a few thousand dollars over a couple weeks on a long horizon task. That's kind of the high level. Claude writes its own harness. There's 6 new workflow patterns that the Anthropic team recommends with this harness on the fly feature. Number 1, classify and act. Task goes into a classifier that routes to different agents. Number 2, fan out and synthesize. Task goes out, it fans out, does stuff, comes back, and then collates it, synthesizes it. Number 3, adversarial verification. Many verifiers and then running a worker. Generate and filter. Generate stuff, get to ideas, filter out with a rubric, deduplicate ideas, and then do a best-of breed amongst them. And discard stuff against your rubric and filtering. Tournament style, where you have attempts to do workflows and then you pair them with judges, which are decision makers. It looks like you can have different tiers of judges here and then a final judge to determine a winner from those attempts at workflows. Finally, number 6, loop until done, where the agent does stuff and there's a checkpoint and it'll say, do you have new findings? No. If no new findings, done. If yes, new findings, spawn another agent, keep going until done. Jonathan, back to you.
So one of the things I did in our codebase is I just tried to talk with Claude Opus 5.0 and say like, hey, I want to, I want to inherit these dynamic workflows. How can I best use them? And we're kind of going and inventorying the type of work that we do in the repository against these workflow patterns and trying to just configure the work environment to like where the next time I do a task that I want it to like trigger on these workflows and use them correctly. So that's the deal. And I think what you're mainly going to focus on is inventory the workflows that you're doing and what's important to what skills have you provided, like as commands for the agent to use? And make sure you're taking the new advice from Claude and simplifying your skills. Leave these like really verbose skills anymore, trim them up, and then leave those as breadcrumbs for your dynamic workflows to pick up on. And then number 3 is rubrics. How do you take the domain-specific knowledge that you've accumulated, the good examples, like what is good in this example, and then synthesize that into a rubric that it can make so it knows it does a good job?
