The old coding-agent pitch was simple: one assistant, one task, one branch, one pull request.
Anthropic is now pushing past that shape.
Claude Opus 4.8 landed on May 28 with the usual model-release language: better benchmarks, better reasoning, same regular price. The more interesting part is what launched beside it. Claude Code now has dynamic workflows, a research-preview feature that lets Claude break a large engineering task into many subtasks, run tens or hundreds of subagents in parallel, verify their work, and return one coordinated result.
That is a different product category.
One agent was never the final form
A single coding agent is useful for local work. Fix this bug. Explain this file. Write this test. Refactor this component.
But real software work is rarely shaped like one clean prompt. A serious migration touches old assumptions, hidden dependencies, build quirks, tests that only fail after the second change, and parts of the system nobody wants to remember. That is why human teams split work across people, review each other, run tests, and argue about the plan before touching production.
Dynamic workflows are Anthropic’s attempt to copy that structure inside the tool.
The launch post says Claude can plan the work, fan it out across subagents, use independent verification, and keep iterating until results converge. Anthropic gives examples like codebase-wide bug hunts, security audits, modernization work, and large migrations. The headline example is a Bun port from Zig to Rust with roughly 750,000 lines of Rust and 99.8% of the existing test suite passing before merge.
That is not autocomplete.
That is automated engineering operations.
The hidden product is trust
Opus 4.8 also adds effort control, cheaper fast mode, higher Claude Code rate limits, and API support for system entries inside the messages array. Taken together, this points at a more mature agent loop: more control over how hard the model thinks, more flexibility in mid-task instructions, more ability to run long asynchronous jobs.
The hard part is not generating code.
The hard part is knowing when the run is safe to trust.
Anthropic says Opus 4.8 is less likely than Opus 4.7 to let flaws in its own code pass without comment. That matters because agent swarms multiply both productivity and failure surface. If one agent hallucinates a migration path, another agent needs to catch it. If one worker makes a local change that breaks an assumption elsewhere, the workflow needs independent pressure from tests, reviewers, and adversarial checks.
This is where coding agents become less like chatbots and more like CI systems with judgment.
The real constraint is cost and blast radius
Anthropic is blunt that dynamic workflows can consume substantially more tokens than normal Claude Code sessions. That is the correct warning. A swarm is expensive. It also deserves stricter boundaries than a one-shot assistant because it can touch more files, run longer, and create more convincing output.
The direction is obvious, though.
Small tasks will stay conversational. Big tasks will become orchestrated. The best agent products will not just answer faster; they will plan, parallelize, verify, and expose enough of the run that a human can decide whether to merge it.
Software teams are not getting one AI coworker.
They are getting a weird little engineering department in a box, and the box had better come with logs.
Sources: Anthropic: Claude Opus 4.8, Claude: Dynamic workflows