/fork runs a background agent carrying your exact system prompt, tools, history, model, and prompt cache, then returns the result to your session. The old /fork behavior (copying the transcript to a new session you control) is now called /branch.
The trigger keyword for launching a dynamic workflow has changed from "workflow" to "ultracode". Saying "use a workflow for this" still works in context, but "ultracode" is now the explicit, unambiguous way to kick one off.
Subagents start fresh with only their initial prompt, system prompt, and tools. A fork is a special kind of subagent that inherits the full conversation history, system prompt, tools, and model, making it essentially a clone of the parent session.
The keyword that activates ultra mode has been updated to 'ultra code.' A team member confirmed the change directly in reply to a user question.
When managing subagents across large contexts, deterministic workflows tend to be a better fit than open-ended agentic approaches for most practical tasks.
Sonnet handles the majority of steps in a typical agentic workflow just fine. Opus isn't always necessary, and the distinction probably deserved more emphasis.
The JS environment gets some determinism treatment, but it's not full temporal durability. A reasonable mental model: each subagent is roughly as durable as hitting /resume, not a persistent long-running process.
Structuring Claude workflows maps closely to multithreaded programming: a serial step queues independent tasks to run in parallel, with minimal cross-dependencies, followed by a serial post-processing step on the results.
An Anthropic team member pushes back on skepticism about agentic workflows, saying the default behavior holds up well in practice, though you can prompt it toward specific patterns if needed.
Looking at a large project now means asking how to decompose it into parallel Claude agents rather than linear steps. The mental model shift is toward maximizing concurrency across independent sub-tasks.
Parallel Claude workflows and remote control have become the biggest shift in one developer's engineering practice in six months. Claude can now handle large, long-running projects that continue making progress while the developer is away from the keyboard.
Even without tight integration, you can ask Claude mid-run how a workflow is going and it will read the transcripts and summarize the state for you.
For fine-grained decisions, one developer still trusts their own judgment over Claude's, though they're considering writing a skill to transfer that knowledge to the model.
The /goal slash command lets Claude run open-ended, autonomous sessions overnight. The main anxiety here is mundane: will the laptop sleep or get interrupted before it finishes.
A new blog post lays out the recommended approach for building data analysis agents: how to structure skills, establish solid data foundations, and set up evaluations that actually tell you if the thing is working.
Before writing a spec, it can make sense to do broad research and exploration across different areas first. The order matters: explore, then specify.
Workflows are being positioned as a significant capability, with a linked article explaining what they are and why they matter.
